Qwen3.5-9B CVE-to-CWE LoRA adapter (run 3, trial 23): weights, tokenizer, model card, predictions

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -1,3 +1,548 @@
 ---
-license: apache-2.0
+# ---- Identity -------------------------------------------------------------
+base_model: Qwen/Qwen3.5-9B
+base_model_relation: adapter
+library_name: peft
+pipeline_tag: text-generation
+language:
+  - en
+license: apache-2.0   # verified: inherited from Qwen/Qwen3.5-9B, whose Hub
+                      # metadata declares license:apache-2.0 and which ships an
+                      # Apache-2.0 LICENSE file (copied here verbatim).
+
+# ---- Discovery ------------------------------------------------------------
+tags:
+  - lora
+  - qlora
+  - sft
+  - trl
+  - peft
+  - text-classification
+  - cve
+  - cwe
+  - vulnerability
+  - security
+
+datasets:
+  - exploitintel/cve-cwe-consensus
+
+metrics:
+  - f1
+  - accuracy
+
+# ---- Structured evaluation ------------------------------------------------
+model-index:
+  - name: qwen3.5-9b-cve-cwe-lora
+    results:
+      - task:
+          type: text-classification
+          name: CVE description to CWE weakness class (117-way, generated as text)
+        dataset:
+          type: exploitintel/cve-cwe-consensus
+          name: exploitintel/cve-cwe-consensus, 300-row natural-distribution sample of the validation split
+          split: validation
+          revision: 606cef101302fc2e0f69fd298f0de28b20a2aacf
+        metrics:
+          - type: f1
+            name: Macro-F1 over the union of gold and predicted classes (81 classes)
+            value: 0.473288
+          - type: f1
+            name: Macro-F1 over gold classes only (68 classes)
+            value: 0.563769
+          - type: accuracy
+            name: Accuracy (= micro-F1 for single-label output)
+            value: 0.71
 ---
+
+# Qwen3.5-9B CVE-to-CWE Classifier (LoRA)
+
+Given the free-text description of a published CVE, emits the single CWE
+identifier that best characterises the root-cause weakness, as one line of
+text (`CWE-79`, `CWE-787`, ...), over the 117-class label space of
+[exploitintel/cve-cwe-consensus](https://huggingface.co/datasets/exploitintel/cve-cwe-consensus).
+
+This is a **LoRA adapter for**
+[Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B), trained with
+**QLoRA (4-bit NF4 base, bf16 compute)** via
+[TRL](https://github.com/huggingface/trl) SFT.
+
+> **The headline number is one draw from a noisy measurement.** This checkpoint
+> scored macro-F1 0.473 on 300 held-out rows, but the identical configuration
+> re-run later in the same search scored 0.395. Read
+> [How these values were chosen](#how-these-values-were-chosen) and
+> [Evaluation](#evaluation) before quoting anything, and do not report 0.473 as
+> the model's expected performance.
+
+## Model details
+
+| | |
+|---|---|
+| Developed by | SASVA AI Model Cognition Labs (MCL) Team |
+| Base model | [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) |
+| Base revision | `c202236235762e1c871ad0ccb60c8ee5ba337b9a` |
+| Base parameters | 9.65B (dense; 9,653,104,368 in the safetensors index, ~19.3 GB bf16) |
+| Architecture family | `qwen3_5` (`Qwen3_5ForConditionalGeneration`, text tower `qwen3_5_text`) |
+| Adaptation | LoRA (`r=128`, `alpha=256`, `dropout=0.15`, rsLoRA off, DoRA off) |
+| Trainable modules | `in_proj_qkv`, `in_proj_z`, `out_proj` (Gated DeltaNet layers); `q_proj`, `k_proj`, `v_proj`, `o_proj` (attention layers); `gate_proj`, `up_proj`, `down_proj` (every layer) |
+| Excluded modules | none (the base is text-only in practice; no adapter tensor touches the vision tower) |
+| Training method | `qlora` (`--load-in-4bit`, run 3 / trial 23) |
+| Refinement | none |
+| Precision | 4-bit NF4 base with double quantisation, bf16 compute; adapter stored in float32 |
+| Language | English |
+| License | Apache-2.0 (inherited from the base model) |
+
+Trainable parameters: **320,864,256** across 200 modules, 3.2975% of the
+9,730,678,000 parameters PEFT counted with the adapter attached. The adapter
+file is 1,283,518,408 bytes (400 tensors, `lora_A` + `lora_B` per module, all
+float32). Every tensor sits under `base_model.model.model.language_model`.
+
+**One Qwen 3.5 structural fact shapes the module list.** Confirmed against the
+base model's `config.json`: the 32 layers follow a 3-linear / 1-full pattern
+(`full_attention_interval: 4`), so 24 layers are Gated DeltaNet linear
+attention and expose `in_proj_qkv` / `in_proj_z` / `out_proj`, while the 8
+full-attention layers (3, 7, 11, ..., 31) are grouped-query attention with 16
+heads over 4 KV heads and expose `q_proj` / `k_proj` / `v_proj` / `o_proj`.
+The adapter's tensor counts match exactly: 48 per linear-attention projection
+(24 layers x A/B), 16 per attention projection (8 layers x A/B), 64 per MLP
+projection (32 layers x A/B). There are no MoE experts in this base.
+
+## Intended use
+
+**Direct use.** Map a CVE description to one CWE id, for triage and
+labelling pipelines that already consume CWE ids. The 117-class label space is
+the dataset's; ids outside it were never seen in training.
+
+The model was trained on a specific prompt shape and that shape is part of the
+contract:
+
+- System prompt (verbatim): *"You are a CVE-to-CWE classifier. Given a CVE
+  vulnerability description, identify the single root-cause CWE weakness class
+  that best characterizes the flaw. Output exactly one CWE identifier (e.g.,
+  CWE-79) on a single line with no explanation or additional text."*
+- User turn: the instruction *"Classify the following CVE description into
+  exactly one CWE weakness class. Reply with the CWE ID only, for example
+  CWE-79."*, a blank line, then the CVE description inside a bare ```` ``` ````
+  fence. This is the exact string the evaluator rendered
+  (`instruction + "\n\n```\n" + description + "\n```"`).
+- Applied through the tokenizer's chat template (`chat_template.jinja`,
+  shipped in this repo) with `add_generation_prompt=True` and
+  `enable_thinking=False`. Do not concatenate strings by hand.
+- The output is one line, `CWE-<n>`. In evaluation all 300 generations were a
+  single well-formed id; the first line of the output is the prediction.
+- Decode greedily (`do_sample=False`) with a small budget; 64 new tokens is
+  what the metric was scored with.
+
+## How to get started
+
+```python
+import torch
+from peft import PeftModel
+from transformers import AutoModelForImageTextToText, AutoTokenizer
+
+BASE = "Qwen/Qwen3.5-9B"
+ADAPTER = "SASVAAI/Qwen-3.5-9B-CVE-to-CWE"
+
+# NOTE: AutoModelForImageTextToText, not AutoModelForCausalLM. Qwen 3.5's
+# CausalLM mapping raises AttributeError: 'Qwen3_5Config' object has no
+# attribute 'vocab_size' on the transformers build used here; the
+# image-text-to-text mapping resolves to Qwen3_5ForConditionalGeneration and
+# loads the text model correctly. Training used the same class.
+tokenizer = AutoTokenizer.from_pretrained(ADAPTER, trust_remote_code=True)
+model = AutoModelForImageTextToText.from_pretrained(
+    BASE, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
+)
+model = PeftModel.from_pretrained(model, ADAPTER)
+model.eval()
+
+# Verbatim from the training invocation. Do not paraphrase.
+SYSTEM = (
+    "You are a CVE-to-CWE classifier. Given a CVE vulnerability description, "
+    "identify the single root-cause CWE weakness class that best characterizes "
+    "the flaw. Output exactly one CWE identifier (e.g., CWE-79) on a single line "
+    "with no explanation or additional text."
+)
+INSTRUCTION = (
+    "Classify the following CVE description into exactly one CWE weakness class. "
+    "Reply with the CWE ID only, for example CWE-79."
+)
+description = (
+    "A stored cross-site scripting flaw in the FAQ page lets an attacker inject "
+    "script that runs in other users' browsers."
+)
+messages = [
+    {"role": "system", "content": SYSTEM},
+    {"role": "user", "content": f"{INSTRUCTION}\n\n```\n{description}\n```"},
+]
+inputs = tokenizer.apply_chat_template(
+    messages, add_generation_prompt=True, enable_thinking=False,
+    return_tensors="pt", return_dict=True,
+).to(model.device)
+
+out = model.generate(**inputs, max_new_tokens=64, do_sample=False)
+print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True).strip())
+# -> CWE-79
+```
+
+The base model is ~19 GB in bfloat16; one 24 GB-class GPU is enough for
+inference with the adapter. Loading the base in 4-bit with `bitsandbytes`
+(`load_in_4bit=True`, `bnb_4bit_quant_type="nf4"`,
+`bnb_4bit_use_double_quant=True`, `bnb_4bit_compute_dtype=torch.bfloat16`)
+reproduces the training-time numerics and needs about 7 GB.
+
+> Decoding matters. The metric was scored greedily with `max_new_tokens=64`
+> through the chat template with thinking disabled. No sampling setting was
+> validated, and enabling thinking changes the prompt the model sees.
+
+## Training details
+
+**Data.** 700 training rows and 300 validation rows built from
+[exploitintel/cve-cwe-consensus](https://huggingface.co/datasets/exploitintel/cve-cwe-consensus)
+at revision `606cef101302fc2e0f69fd298f0de28b20a2aacf` by a deterministic
+script (`autocatalyst.datagen.cve_cwe_consensus`, seed 0). No model generated
+any training content.
+
+Row selection, as recorded in the builder's manifest:
+
+- Source rows are `{"messages": [system, user, assistant]}`; the user turn is
+  the CVE description, the assistant turn the label.
+- Rows whose label is not exactly one `CWE-<n>` id were dropped (about 9% of
+  the dataset are comma-separated multi-label lists). Rows whose id is outside
+  the dataset's 117-class set, or whose description is empty, were dropped.
+  Duplicate descriptions within a split were collapsed to the first.
+- **Train**: from the dataset's `train` split (50,074 raw rows, 44,337 kept
+  after dropping 4,390 multi-label or malformed and 1,347 duplicate rows),
+  700 rows were sampled with one row guaranteed per class and the remainder
+  weighted by the square root of class frequency, so the head classes do not
+  crowd out the tail that macro-F1 scores. All 117 classes are present.
+- **Validation**: from the dataset's own `validation` split (11,052 raw rows,
+  8,878 kept), 300 rows sampled at the natural class distribution. 68 classes
+  appear; every one of them is in the training set; no description text is
+  shared with the 700 training rows (0 overlaps measured).
+
+| | |
+|---|---|
+| Train samples | 700 rows / 117 classes |
+| Validation samples | 300 rows / 68 classes |
+| Text overlap | 0 rows |
+| Prompt format | chat template + system prompt + instruction/fenced-description user turn (see Intended use) |
+| Loss masking | answer tokens only; prompt tokens set to `-100` |
+| Truncation | prompt left-truncated to fit `max_seq_len` 2048; answer never truncated |
+
+The dataset's head is heavy: in the 300 validation rows CWE-79 appears 49
+times, CWE-862 22, CWE-89 17, CWE-200 and CWE-22 13 each, and 68 classes share
+the rest.
+
+An LLM (Claude Opus 4.6, `claude-opus-4-6`, via an internal inference gateway)
+proposed the hyperparameters the search tried. It generated no training
+content and computed no metric.
+
+### Method
+
+| | |
+|---|---|
+| SFT method | `qlora` |
+| Base quantisation during training | 4-bit NF4, double quantisation, bf16 compute (`bitsandbytes`) |
+| Refinement stage | none |
+| Attention implementation | `flash_attention_2` for the 8 GQA layers; the 24 Gated DeltaNet layers ran the PyTorch fallback because `flash-linear-attention` was not installed |
+| Auto class | `AutoModelForImageTextToText` (see How to get started) |
+| Hardware | 8x NVIDIA H100 80GB HBM3, `torchrun --nproc_per_node=8` |
+
+The project allowed two methods for this run (`bf16_lora`, `qlora`) and the
+search tried both; see the trial table below.
+
+No refinement stage ran; the published weights are the SFT adapter.
+
+### Final hyperparameters
+
+| Hyperparameter | Value | Source |
+|---|---|---|
+| `learning_rate` | 0.0002 | `[TRAIN]` cmdline |
+| `lr_scheduler_type` | cosine | `[TRAIN]` cmdline |
+| `num_train_epochs` | 5 | `[TRAIN]` cmdline, `trainer_state.json` |
+| `per_device_train_batch_size` | 2 | `[TRAIN]` cmdline, `trainer_state.json` |
+| `gradient_accumulation_steps` | 4 | `[TRAIN]` cmdline |
+| `max_seq_length` | 2048 | `[TRAIN]` cmdline |
+| `warmup_ratio` | 0.05 | `[TRAIN]` cmdline |
+| `weight_decay` | 0.05 | `[TRAIN]` cmdline |
+| `loraplus_lr_ratio` | 2.0 (LoRA `B` matrices at 2x the learning rate) | `[TRAIN]` cmdline |
+| `lora_r` / `lora_alpha` / `lora_dropout` | 128 / 256 / 0.15 | `adapter_config.json` |
+| `use_rslora` / `use_dora` | `false` / `false` | `adapter_config.json` |
+| `target_modules` | the 10 listed in Model details | `adapter_config.json` |
+| `load_in_4bit` | `true` (NF4, double quant, bf16 compute) | `[TRAIN]` cmdline |
+
+**Effective batch size: 64** (`2 x 4 x 8`). Optimizer steps: 55 (11 per
+epoch).
+
+`neftune_noise_alpha` (0.0), `use_liger_kernel`, `use_sample_packing` and
+`lora_init` (`default`) were left at their no-op defaults. KD parameters are
+omitted deliberately: this is a `qlora` run, not a distillation run.
+
+> Provenance note: every value above was recovered from the platform database
+> (`runs`, `experiments`, `events` tables for run 3) and cross-checked against
+> the literal `[TRAIN]` command line recorded in the run log and against the
+> shipped `adapter_config.json`.
+
+### How these values were chosen
+
+> These hyperparameters were selected by an automated search
+> (`autocatalyst.cli.run_autoresearch`): an agent proposes one change at a
+> time, runs train then eval, and keeps or discards on `f1_macro` (higher is
+> better).
+
+Run 3 ran **46 trials in 7 h 17 m** (2026-09-08 20:29 to 2026-09-09 03:46);
+35 scored and 11 errored. This checkpoint is **trial 23**, the run's best.
+
+**Every trial is comparable.** The search's five "data strategies" all
+re-adapted the same fixed 700-row training file (the `train_v1` ... `train_v5`
+files below hold the same 700 rows) and every trial was scored on the same 300
+validation rows, so the whole table is one comparison.
+
+| # | Method | rsLoRA | r | Dropout | LR | Epochs | Grad accum | Warmup | Weight decay | LoRA+ ratio | Data | f1_macro | Kept |
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| 1 | bf16_lora | no | 16 | 0.05 | 2e-4 | 2 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.271028 | yes |
+| 2 | qlora | no | 16 | 0.05 | 2e-4 | 2 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.323589 | yes |
+| 3 | bf16_lora | no | 16 | 0.05 | 2e-4 | 2 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.271161 | no |
+| 4 | qlora | no | 16 | 0.05 | 1e-4 | 2 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.224824 | no |
+| 5 | qlora | no | 16 | 0.05 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.328347 | yes |
+| 6 | qlora | no | 16 | 0.05 | 2e-4 | 8 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.315655 | no |
+| 7 | qlora | no | 32 | 0.05 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.339796 | yes |
+| 8 | qlora | no | 32 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.358649 | yes |
+| 9 | qlora | no | 32 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v1 | 0.379137 | yes |
+| 10 | qlora | no | 64 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v1 | 0.411651 | yes |
+| 11 | qlora | yes | 64 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | 0.385129 | no |
+| 12 | qlora | no | 64 | 0.05 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v1 | 0.405714 | no |
+| 13 | qlora | no | 64 | 0.15 | 2e-4 | 5 | 8 | 0.05 | 0.01 | 2.0 | train_v1 | 0.377237 | no |
+| 14 | qlora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v1 | 0.440878 | yes |
+| 15 | qlora | no | 128 | 0.15 | 2e-4 | 8 | 4 | 0.05 | 0.01 | 2.0 | train_v1 | 0.377763 | no |
+| 16 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | error | no |
+| 17 | qlora | no | 128 | 0.15 | 1.5e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v1 | 0.386976 | no |
+| 18 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 1.0 | train_v1 | error | no |
+| 19 | qlora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v1 | 0.396817 | no |
+| 20 | qlora | no | 128 | 0.15 | 2e-4 | 8 | 4 | 0.05 | 0.01 | 2.0 | train_v2 | 0.407741 | no |
+| 21 | bf16_lora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v2 | 0.403981 | no |
+| 22 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 1.0 | train_v2 | error | no |
+| **23** | **qlora** | **no** | **128** | **0.15** | **2e-4** | **5** | **4** | **0.05** | **0.05** | **2.0** | **train_v2** | **0.473288** | **yes** |
+| 24 | qlora | no | 128 | 0.10 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v2 | 0.444999 | no |
+| 25 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v2 | error | no |
+| 26 | qlora | no | 128 | 0.15 | 2.5e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v2 | 0.427152 | no |
+| 27 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v2 | error | no |
+| 28 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v2 | error | no |
+| 29 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v3 | error | no |
+| 30 | qlora | no | 128 | 0.15 | 2e-4 | 8 | 4 | 0.05 | 0.05 | 2.0 | train_v3 | 0.384342 | no |
+| 31 | bf16_lora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v3 | 0.403688 | no |
+| 32 | qlora | no | 128 | 0.05 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v3 | 0.415561 | no |
+| 33 | qlora | no | 128 | 0.15 | 1.75e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v3 | 0.435022 | no |
+| 34 | qlora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.03 | 0.05 | 2.0 | train_v3 | 0.416758 | no |
+| 35 | qlora | no | 128 | 0.15 | 2e-4 | 8 | 4 | 0.05 | 0.05 | 2.0 | train_v4 | 0.406259 | no |
+| 36 | bf16_lora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v4 | 0.408250 | no |
+| 37 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v4 | error | no |
+| 38 | qlora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v4 | 0.395110 | no |
+| 39 | qlora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.01 | 2.0 | train_v4 | 0.437202 | no |
+| 40 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v4 | error | no |
+| 41 | qlora | no | 128 | 0.15 | 2e-4 | 8 | 4 | 0.05 | 0.05 | 2.0 | train_v5 | 0.391961 | no |
+| 42 | bf16_lora | no | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v5 | 0.428970 | no |
+| 43 | qlora | no | 128 | 0.05 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 2.0 | train_v5 | 0.423615 | no |
+| 44 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v5 | error | no |
+| 45 | qlora | yes | 128 | 0.15 | 2e-4 | 5 | 4 | 0.05 | 0.05 | 1.0 | train_v5 | error | no |
+| 46 | qlora | no | 128 | 0.15 | 2e-4 | 5 | 8 | 0.05 | 0.05 | 2.0 | train_v5 | 0.402578 | no |
+
+All trials: `lora_alpha = 2 x r`, cosine schedule, batch 2 per device,
+`max_seq_len` 2048.
+
+**What the search actually established.**
+
+- **Rank dominates, up to a point.** Holding QLoRA, 5 epochs, dropout 0.15,
+  LR 2e-4, weight decay 0.01, LoRA+ 2.0: `r` 32 → 64 → 128 moved macro-F1
+  0.379137 → 0.411651 → 0.440878 (#9, #10, #14).
+- **LoRA+ helped.** The one clean A/B at `r=32` (#8 → #9) moved 0.358649 →
+  0.379137 by putting the `B` matrices at 2x the learning rate.
+- **More epochs hurt.** 8 epochs lost to 5 every time it was tried at `r=128`
+  (#15, #20, #30, #35, #41), by 0.01 to 0.06.
+- **4-bit versus bf16 base: no measurable difference.** At `r=16` the 4-bit
+  base beat bf16 (#2 at 0.323589 against #1 and #3 at 0.271); at `r=128` with
+  weight decay 0.05 the three bf16 runs (#31, #36, #42: 0.404 to 0.429) sit
+  inside the two 4-bit runs of the same configuration (#38 at 0.395, #23 at
+  0.473). Quantising the base cost nothing this search could detect.
+- **rsLoRA at `r=128` diverges.** All 11 rsLoRA trials at `r=128` (#16, #18,
+  #22, #25, #27, #28, #29, #37, #40, #44, #45) were aborted by the training
+  watchdog at the third logging step with gradient norm above 10 (the
+  configured limit; observed 12.3 in the first). The likely cause is the
+  scaling rule: rsLoRA scales updates by `alpha / sqrt(r)` instead of
+  `alpha / r`, which with `alpha = 2r` is `2 sqrt(r)`, about 22.6 at `r=128`
+  against 2.0 without it, an 11x larger effective update. The one rsLoRA
+  trial that finished (#11, `r=64`) scored below its non-rsLoRA sibling (#10).
+  The 11 aborted trials are why the search shows 46 trials but 35 scores.
+
+**What it did not establish: the winning margin.** The search's decision
+"weight decay 0.01 → 0.05 improved 0.440878 → 0.473288" (#14 → #23) does not
+survive the re-runs. Identical configurations were run more than once because
+each "data strategy" restarted the inner loop on the same data:
+
+| Configuration (all QLoRA, r=128, dropout 0.15, LR 2e-4, 5 epochs, grad accum 4, warmup 0.05, LoRA+ 2.0) | Trials | f1_macro | Spread |
+|---|---|---|---|
+| weight decay 0.05 (**this checkpoint's config**) | #23, #38 | 0.473288, 0.395110 | 0.078 |
+| weight decay 0.01 | #14, #19, #39 | 0.440878, 0.396817, 0.437202 | 0.044 |
+| weight decay 0.05, bf16 base instead of 4-bit | #31, #36, #42 | 0.403688, 0.408250, 0.428970 | 0.025 |
+| weight decay 0.05, 8 epochs | #30, #35, #41 | 0.384342, 0.406259, 0.391961 | 0.022 |
+| weight decay 0.05, dropout 0.05 | #32, #43 | 0.415561, 0.423615 | 0.008 |
+
+Seeds were not pinned, so these differ only in initialisation and data order.
+The 0.078 gap between #23 and #38 is larger than the 0.032 "improvement" the
+search kept, and larger than most differences in the table. Read 0.473 as the
+high draw of a configuration whose expected score is somewhere in the low-to-mid
+0.4s, and treat any two trials within about 0.05 of each other as tied.
+
+**Search space.** Six knobs were varied (`LORA_R`, `LORA_DROPOUT`,
+`LEARNING_RATE`, `EPOCHS`, `WEIGHT_DECAY`, `LORAPLUS_LR_RATIO`) plus one
+`GRAD_ACCUM` probe (#13, #46), one warmup probe (#34), the bf16-vs-4-bit
+method switch, and the rsLoRA attempts. `LR_SCHEDULER`, `MAX_SEQ_LEN`,
+`BATCH_SIZE`, `NEFTUNE_NOISE_ALPHA`, `USE_DORA`, `LORA_INIT`,
+`USE_LIGER_KERNEL` and `USE_SAMPLE_PACKING` were never moved.
+
+**Observed training metrics** (this checkpoint).
+
+| | |
+|---|---|
+| Final train loss (mean over the run) | 0.6052066683769226 |
+| Final eval loss (teacher-forced, answer tokens) | 0.9717274904251099 |
+| Eval mean token accuracy | 0.8136 |
+| Train runtime | 387.7625 s |
+| Total FLOPs | 4.311845712166912e+16 |
+| Throughput | 9.026 samples/s, 0.142 steps/s |
+
+55 optimizer steps ran. The logged train loss fell from 1.2962 at step 10 (grad
+norm 1.06) to the run mean of 0.6052; the reported train loss is the mean over
+the run, not a converged value.
+
+## Evaluation
+
+**Protocol.** All 300 validation rows, greedy decoding, `max_new_tokens=64`,
+prompts rendered through the chat template with `enable_thinking=False`. The
+project's `classification` evaluator takes the first line of the generation as
+the predicted label and compares it to the gold id. Generation ran through
+Hugging Face `generate` in batches of 4 (`max_input_len` 4096) after vLLM
+0.19.1 refused the LoRA on this architecture and the evaluator fell back; the
+whole pass took 63 s on 8 GPUs. Every one of the 300 outputs was a single
+well-formed `CWE-<n>` id, so no prediction was lost to formatting, and the
+64-token budget cut nothing (the longest output is one id).
+
+| Metric | Value |
+|---|---|
+| Macro-F1, union of gold and predicted classes (**the search metric**) | 0.473288 |
+| Macro-F1, gold classes only | 0.563769 |
+| Accuracy (= micro-F1) | 0.71 (213 / 300) |
+| Classes in gold / predicted / union | 68 / 72 / 81 |
+
+**Two macro-F1 numbers, one convention.** The evaluator averages F1 over the
+*union* of gold and predicted classes, so the 13 classes the model predicted
+that never occur in the 300 gold rows each contribute an F1 of 0 and pull the
+macro average down from 0.5638 to 0.4733. Both are reported; the union
+convention is the one the search optimised and the one the published baselines
+below appear to use, but check before comparing.
+
+**Published comparison points.** The dataset authors report, on their own
+evaluation of the same dataset (a much larger split than these 300 rows):
+
+| Model | Method | Micro-F1 | Macro-F1 |
+|---|---|---|---|
+| Qwen3-32B ([exploitintel/cve-cwe-qwen3-32b](https://huggingface.co/exploitintel/cve-cwe-qwen3-32b)) | QLoRA r=16 | 0.729 | 0.595 |
+| Qwen3-8B (same author, same recipe) | QLoRA r=16 | 0.702 | 0.511 |
+| **This adapter** (Qwen3.5-9B, QLoRA r=128, 700 training rows) | | **0.71** | **0.473** |
+
+Those models trained on the full ~44K-row training split; this adapter saw
+700 rows. The micro-F1 is in the same range; the macro-F1 is 0.04 to 0.12
+lower, which is the long tail this adapter had one to twelve examples of per
+class to learn from. The evaluation sets also differ in size and composition,
+so read this as context, not a controlled comparison.
+
+**Sibling run on the same 700 / 300 split.** A Gemma 4 E4B adapter trained by
+the same platform on this split reached accuracy 0.697 and union macro-F1
+0.4587, essentially tied with this checkpoint given the spread above.
+
+**Baseline for comparison. Not measured.** The untuned `Qwen/Qwen3.5-9B` was
+never scored on these 300 rows, so nothing here quantifies how much of the
+score the fine-tuning is responsible for. This is the most important gap in
+this card.
+
+**This is a validation split the search selected against.** 35 trials were
+scored on these same 300 rows and the best was kept, so expect optimistic bias
+on top of the re-run spread already described. The rows are drawn from the
+dataset authors' own validation split, so they are unseen CVEs from the same
+period as training, not future CVEs.
+
+**The evaluation set is reproducible.** `predictions.jsonl` in this repo holds
+every one of the 300 rows: instruction, description, prediction and gold. The
+builder's manifest (dataset revision, seed, drop counts, per-class counts) is
+summarised under Training details.
+
+## Limitations and bias
+
+**One number, wide error bars.** The same configuration scored 0.473 and 0.395
+in two runs. Anyone deploying this should re-evaluate on their own data rather
+than trust either figure.
+
+**No baseline, so no established gain.** See Evaluation.
+
+**Head classes dominate what accuracy measures.** CWE-79 alone is 16% of the
+validation rows. A model that got only the top ten classes right would post a
+respectable accuracy and a poor macro-F1; the two numbers here disagree by
+0.24 for that reason.
+
+**Tail classes were barely trained.** 117 classes over 700 rows means many
+classes had a single training example. Expect the model to fall back to a
+frequent neighbour (CWE-20 for input validation issues, CWE-200 for disclosure)
+when the description is ambiguous, and to emit ids it saw rarely with low
+reliability. It also produced 13 ids in evaluation that never occur in the 300
+gold rows; some may be reasonable alternative labels, some are wrong.
+
+**Only one label.** Real CVEs often carry two CWE ids (about 9% of the source
+dataset). The training data dropped those rows, so the model always commits to
+one.
+
+**Prompt shape is the contract.** Change the system prompt, the instruction
+sentence, the code fence, or enable thinking, and you are evaluating a model
+nobody measured.
+
+**Domain narrowness.** English CVE descriptions in NVD / CNA style. Advisories
+in other formats, other languages, source-code inputs and exploit write-ups are
+unmeasured.
+
+**Inherits all biases and limitations of the base model.** This adapter changes
+3.3% of the parameters and was not evaluated for safety or fairness. The base
+model's own card governs those properties.
+
+## Environmental impact
+
+| | |
+|---|---|
+| Hardware | 8x NVIDIA H100 80GB HBM3 |
+| Training time | 6.46 minutes (387.7625 s) |
+| Cloud provider / region | on-premise |
+
+Covers this trial only. The full 46-trial search that selected it took 7 h 17 m
+on the same hardware.
+
+## Framework versions
+
+- PEFT 0.18.1
+- TRL: 1.0.0
+- Transformers: 5.7.0.dev0 (git main)
+- Pytorch: 2.5.1+cu121
+- bitsandbytes: 0.49.2
+- flash-attn: 2.8.3
+- Python: 3.12
+
+PEFT's version is the one recorded in `adapter_config.json` at save time; the
+rest are the pinned versions of the training environment. `transformers` is a
+git-main build: the `qwen3_5` architecture is not in the stable PyPI release.
+
+## Citation
+
+```bibtex
+@misc{qwen35_cve_cwe_lora_2026,
+  title  = {Qwen3.5-9B CVE-to-CWE Classifier LoRA},
+  author = {Banerjee, Aaron and Anbuselvan, Pooja and Jodhpurkar, Om},
+  year   = {2026},
+  url    = {https://huggingface.co/SASVAAI/Qwen-3.5-9B-CVE-to-CWE}
+}
+```

adapter_config.json

@@ -0,0 +1,49 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3.5-9B",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 256,
+  "lora_bias": false,
+  "lora_dropout": 0.15,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 128,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "in_proj_z",
+    "v_proj",
+    "o_proj",
+    "q_proj",
+    "k_proj",
+    "down_proj",
+    "up_proj",
+    "out_proj",
+    "in_proj_qkv"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ab2a7cf8c629725381aa12f056fd3e266f52d8bd8efce6c132711dcbe014663
+size 1283518408

all_results.json

@@ -0,0 +1,7 @@
+{
+    "total_flos": 4.311845712166912e+16,
+    "train_loss": 0.6052066683769226,
+    "train_runtime": 387.7625,
+    "train_samples_per_second": 9.026,
+    "train_steps_per_second": 0.142
+}

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

eval_results.json

@@ -0,0 +1,8 @@
+{
+  "metrics": {
+    "accuracy": 0.71,
+    "f1_micro": 0.71,
+    "f1_macro": 0.47328777778730374
+  },
+  "num_samples": 300
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

tokenizer_config.json

@@ -0,0 +1,32 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}

train_results.json

@@ -0,0 +1,7 @@
+{
+    "total_flos": 4.311845712166912e+16,
+    "train_loss": 0.6052066683769226,
+    "train_runtime": 387.7625,
+    "train_samples_per_second": 9.026,
+    "train_steps_per_second": 0.142
+}

trainer_state.json

@@ -0,0 +1,104 @@
+{
+  "best_global_step": 55,
+  "best_metric": 0.9717274904251099,
+  "best_model_checkpoint": "checkpoint-55",
+  "epoch": 5.0,
+  "eval_steps": 9999,
+  "global_step": 55,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "entropy": 1.1605470314621926,
+      "epoch": 0.9090909090909091,
+      "grad_norm": 1.0625,
+      "learning_rate": 0.0001935016242685415,
+      "loss": 1.29622802734375,
+      "mean_token_accuracy": 0.7211568385362626,
+      "num_tokens": 130008.0,
+      "step": 10
+    },
+    {
+      "entropy": 0.7875008955597878,
+      "epoch": 1.8181818181818183,
+      "grad_norm": 0.328125,
+      "learning_rate": 0.00015680647467311557,
+      "loss": 0.7808365821838379,
+      "mean_token_accuracy": 0.8202625662088394,
+      "num_tokens": 260150.0,
+      "step": 20
+    },
+    {
+      "entropy": 0.5799194134771823,
+      "epoch": 2.7272727272727275,
+      "grad_norm": 0.455078125,
+      "learning_rate": 0.0001,
+      "loss": 0.5477523803710938,
+      "mean_token_accuracy": 0.8625506401062012,
+      "num_tokens": 391211.0,
+      "step": 30
+    },
+    {
+      "entropy": 0.3913830861449242,
+      "epoch": 3.6363636363636362,
+      "grad_norm": 0.3984375,
+      "learning_rate": 4.3193525326884435e-05,
+      "loss": 0.35632944107055664,
+      "mean_token_accuracy": 0.9054566085338592,
+      "num_tokens": 519150.0,
+      "step": 40
+    },
+    {
+      "entropy": 0.3066374149173498,
+      "epoch": 4.545454545454545,
+      "grad_norm": 0.4296875,
+      "learning_rate": 6.498375731458528e-06,
+      "loss": 0.24866189956665039,
+      "mean_token_accuracy": 0.9313088282942772,
+      "num_tokens": 649203.0,
+      "step": 50
+    },
+    {
+      "epoch": 5.0,
+      "eval_entropy": 0.5388994969819721,
+      "eval_loss": 0.9717274904251099,
+      "eval_mean_token_accuracy": 0.8135968666327628,
+      "eval_num_tokens": 714955.0,
+      "eval_runtime": 6.6266,
+      "eval_samples_per_second": 45.272,
+      "eval_steps_per_second": 2.867,
+      "step": 55
+    },
+    {
+      "epoch": 5.0,
+      "step": 55,
+      "total_flos": 4.311845712166912e+16,
+      "train_loss": 0.6052066683769226,
+      "train_runtime": 387.7625,
+      "train_samples_per_second": 9.026,
+      "train_steps_per_second": 0.142
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 55,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 5,
+  "save_steps": 9999,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": true
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 4.311845712166912e+16,
+  "train_batch_size": 2,
+  "trial_name": null,
+  "trial_params": null
+}