Upload validated TinyCeNN standalone release

README.md

@@ -1,115 +1,49 @@
 ---
 library_name: transformers
-license: apache-2.0
-base_model: Qwen/Qwen3.5-0.8B
+pipeline_tag: text-generation
 tags:
-- qwen3.5
 - tinycenn
-- recurrent-attention
-- efficient-attention
-- standalone
-- safetensors
+- cenn
+- language-modeling
+- text-generation
+- research
 ---
 
 # Qwen3.5-0.8B-CeNN-Integrated-V1-Standalone
 
-Standalone release of `vtava/Qwen3.5-0.8B-CeNN-Integrated-V1` based on **Qwen3.5-0.8B**.
-
-## Why this release is different
-
-This repository contains the **complete model weights**, tokenizer, Qwen3.5 config,
-and the minimal TinyCeNN inference runtime required for this architecture. You do
-not need to clone the TinyCeNN-LM repository.
-
-The release was uploaded only after a fresh Python process reconstructed the
-custom architecture, loaded `model.safetensors` with strict key checking, and
-produced finite logits with the same validation next token as the source model.
+Research artifact from **TinyCeNN-LM**. Architecture: `TinyCeNN-LM experiment`.
 
 ## Architecture
 
-- Variant: `cenn_integrated`
-- Custom class: `Qwen35IntegratedAttention`
-- Replaced full-attention layers: `[3, 23]`
-- Parameters: `753,609,552`
-- Custom TinyCeNN tensors: `10`
-- Standalone reload: **PASS**
-- Probe next token: ` Vienna`
-
-## Install
-
-Until Qwen3.5 support is available in a stable Transformers release used by your
-environment:
-
-```bash
-pip install torch safetensors huggingface_hub
-pip install git+https://github.com/huggingface/transformers.git@main
-```
-
-## Load
+- Architecture/run type: `TinyCeNN-LM experiment`
+- Base model: `not recorded`
+- Dataset: `Not recorded`
+- Source code: https://github.com/vtavakkoli/TinyCeNN-LM
 
-```python
-from huggingface_hub import snapshot_download
-from pathlib import Path
-import sys
+## Latest saved results
 
-path = Path(snapshot_download("vtava/Qwen3.5-0.8B-CeNN-Integrated-V1-Standalone"))
-sys.path.insert(0, str(path))
-from load_model import load_model
-model, tokenizer = load_model(path, device="cuda")
-```
-
-## TinyCeNN QuickCheck v1
-
-This is a small deterministic regression/sanity suite designed to finish quickly.
-It is **not** an official MMLU, GPQA, IFEval, MMMLU, or LongBench run and its
-percentage must not be compared directly with those benchmark scores.
+| Metric | Value |
+|---|---:|
+| `feature_dim` | 64 |
 
-**Score: 80.0% (4/5)**
+The Hugging Face repository keeps timestamped run artifacts under `runs/`. This preserves training reports, configs and run metadata independently of the temporary Colab filesystem.
 
-| Category | Result | Prediction | Expected |
-|---|---:|---:|---:|
-| Knowledge | PASS | B | B |
-| STEM | PASS | C | C |
-| Reasoning | FAIL | C | D |
-| Multilingual | PASS | C | C |
-| Context | PASS | B | B |
+## Saved experiment files
 
-Full machine-readable results are in `quickcheck.json`.
+- `config.json`
+- `generation_config.json`
+- `release_report.json`
+- `standalone_config.json`
+- `tokenizer_config.json`
 
-## Qwen3.5-0.8B reference context
+## Reproducibility
 
-The values below are reference values supplied for context; this release script
-does **not** rerun those benchmark suites.
+Run the matching notebook from the TinyCeNN-LM repository. Colab notebooks use a Hugging Face write token from the `HF_TOKEN` Colab Secret; tokens should never be pasted into notebook source.
 
-| Benchmark | Qwen3.5-0.8B |
-|---|---:|
-| MMLU-Pro (non-thinking) | 29.7 |
-| MMLU-Redux (non-thinking) | 48.5 |
-| C-Eval (non-thinking) | 46.4 |
-| IFEval (non-thinking) | 52.1 |
-| MMMLU (non-thinking) | 34.1 |
-| MMLU-Pro (thinking) | 42.3 |
-| GPQA (thinking) | 11.9 |
-| LongBench v2 (thinking) | 26.1 |
-| MMMLU (thinking) | 44.3 |
-| Global PIQA (thinking) | 59.4 |
-| WMT24++ (thinking) | 27.2 |
-
-## Files
+## Limitations
 
-- `model.safetensors` — complete reconstructed model weights
-- `config.json` — Qwen3.5 model configuration
-- tokenizer files
-- `tinycenn_lm/` — minimal inference-only TinyCeNN runtime
-- `load_model.py` — standalone loader with strict weight checking
-- `standalone_config.json` — custom architecture metadata
-- `quickcheck.json` — fast sanity-suite results
-- `validation_probe.json` — reload-equivalence probe
-- `release_report.json` — release summary
+This is a research checkpoint. Metrics saved here are the metrics produced by the corresponding training notebook/script; unless explicitly marked as held-out evaluation, they should not be treated as publication-grade benchmark results. Generation quality can differ substantially from the base model.
 
-## Limitations
+## Citation
 
-- Research model; not a production inference kernel.
-- The bundled runtime is reference PyTorch code, not a fused CUDA kernel.
-- QuickCheck is only a regression test, not a publication-grade benchmark.
-- GGUF still requires native TinyCeNN operator support in llama.cpp.
+If you use this experimental checkpoint, cite the TinyCeNN-LM repository and the upstream base model.

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b601d5794f55ceb4f8cdfc2e621defa235da3226d129fdc18b9583077812765a
+oid sha256:3d66e5c849ac49974a41444662995ca4fe23c6ef6b07bdcbaeb2fbe74cb5193d
 size 1509689856

run_manifest.json

@@ -0,0 +1,18 @@
+{
+  "run_id": "vtava__Qwen3.5-0.8B-CeNN-20260917T221342Z",
+  "created_utc": "2026-09-17T22:13:42.455764+00:00",
+  "repo_id": "vtava/Qwen3.5-0.8B-CeNN-Integrated-V1-Standalone",
+  "notebook": null,
+  "python": "3.13.15",
+  "platform": "Linux-6.6.122+-x86_64-with-glibc2.39",
+  "reports": [
+    "config.json",
+    "generation_config.json",
+    "release_report.json",
+    "standalone_config.json",
+    "tokenizer_config.json"
+  ],
+  "torch": "2.11.0+cu128",
+  "cuda_available": true,
+  "gpu": "Tesla T4"
+}