Upload folder using huggingface_hub

README.md

@@ -0,0 +1,95 @@
+---
+base_model:
+- ModalityDance/latent-tts-coconut
+license: mit
+pipeline_tag: text-generation
+library_name: transformers
+tags:
+- latent-reasoning
+- continuous-thought
+- coconut
+- grpo
+- reinforcement-learning
+datasets:
+- gsm8k
+---
+
+# SVP-V-GRPO · COCONUT GPT-2
+
+A [COCONUT](https://huggingface.co/ModalityDance/latent-tts-coconut) GPT-2 (124M)
+latent-reasoning model post-trained with **SVP-V-GRPO** — reinforcement learning
+whose exploration comes from *weight-space* perturbation of the attention value
+projections rather than from token sampling.
+
+Only the value-projection weights (`W_V`, the V columns of each `c_attn`) differ
+from the base checkpoint; everything else is untouched. Inference is ordinary
+greedy decoding — the perturbation is a training-time exploration mechanism and
+is **not** used at deployment.
+
+Part of the **SVP-V** family: `svp-v-coconut-gpt2` (this model) and
+`svp-v-coconut-llama1b` (LLaMA-3.2-1B, in training).
+
+## Results
+
+Clean greedy decoding, `max_new_tokens=64` (six latent steps + marker + answer),
+canonical answer extraction. All rows measured in one harness.
+
+| Model (GPT-2 124M) | GSM8K | GSM-Hard | SVAMP | ASDiv-A | MultiArith | GSM-Plus |
+|---|---|---|---|---|---|---|
+| COCONUT (base) | 34.1 | 7.7 | 35.6 | 60.2 | 80.9 | 17.4 |
+| SLPO-COCONUT | 34.9 | 7.6 | 34.3 | 58.7 | 82.8 | 18.2 |
+| SIM-CoT (COCONUT) | 44.7 | 9.3 | 40.6 | 67.2 | 90.5 | 21.5 |
+| SIM-CoT (CODI) | 39.3 | 8.7 | 40.1 | 64.3 | 91.0 | 22.6 |
+| CODI | 42.5 | 9.3 | 40.0 | 65.4 | 91.9 | 23.1 |
+| SLPO-CODI | 42.9 | 9.5 | 43.0 | 67.5 | 90.5 | 24.1 |
+| **This model** | **50.5** | **11.2** | **45.2** | **72.5** | **94.1** | **28.3** |
+
+GSM8K coverage under the 16-sample deployment ensemble rises alongside single-shot
+accuracy (63.2 → 70.7 over training), i.e. RL here does not collapse the sampling
+diversity it was trained under.
+
+## Usage
+
+This is a fixed-length continuous-thought model: the prompt ends with
+`<|start-latent|>`, six latent steps feed each step's last hidden state back as
+the next input embedding, `<|end-latent|>` closes the latent phase, and the answer
+is then decoded as ordinary tokens. A plain `generate()` call will **not**
+reproduce the numbers above — the two-phase loop is required.
+
+```python
+from transformers import AutoTokenizer, GPT2LMHeadModel
+
+REPO = "jihwan1205/svp-v-coconut-gpt2"
+tok = AutoTokenizer.from_pretrained(REPO)
+model = GPT2LMHeadModel.from_pretrained(REPO).eval().cuda()
+# then run the two-phase latent loop (6 latent steps, then decode)
+```
+
+Reference implementation of the loop: `generate_batch` in
+`src/generation/engine.py` of the SVP repository, or the original COCONUT
+inference code.
+
+## Training
+
+| | |
+|---|---|
+| Base | `ModalityDance/latent-tts-coconut` (COCONUT GPT-2, 124M) |
+| Data | GSM8K-Aug training stream (385k), 9 epochs |
+| Algorithm | GRPO — G=32 rollouts/prompt, B=8 prompts/update, μ=2 inner epochs, DAPO mixed-outcome filter, Dr.GRPO advantage, k3 KL (β=0.02) to the frozen base |
+| Exploration | SVP: per-rollout multiplicative Gaussian jitter on the singular values of `W_V` (σᵢ → σᵢ(1+αgᵢ), α=0.6), re-anchored every 50 iterations |
+| Trained parameters | V columns of every `c_attn` (≈ 7M of 124M) |
+| Optimizer | AdamW, lr 3e-5 constant, grad-clip 1.0 |
+| Reward | Final-answer correctness only |
+
+## Limitations
+
+- English grade-school arithmetic word problems only; the four transfer benchmarks
+  above (SVAMP, ASDiv-A, MultiArith, GSM-Plus) are the extent of tested generalization.
+- GSM-Hard remains low (11.2) — large-magnitude arithmetic is a limit of the
+  124M backbone, not something RL fixed.
+- Requires the COCONUT two-phase inference loop; it is not a drop-in chat model.
+
+## License
+
+MIT, inherited from the base checkpoint (which derives from
+`openai-community/gpt2`).

added_tokens.json

@@ -0,0 +1,5 @@
+{
+  "<|end-latent|>": 50258,
+  "<|latent|>": 50259,
+  "<|start-latent|>": 50257
+}

config.json

@@ -0,0 +1,45 @@
+{
+  "activation_function": "gelu_new",
+  "add_cross_attention": false,
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.1,
+  "bos_token_id": 50256,
+  "dtype": "float32",
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "initializer_range": 0.02,
+  "latent_end_id": -100,
+  "latent_id": -100,
+  "latent_start_id": -100,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "pad_token_id": 50256,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": false,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "target_id": -100,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.13.0",
+  "use_cache": true,
+  "vocab_size": 50260
+}

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "pad_token_id": 50256,
+  "transformers_version": "5.13.0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8e113a71f6ba0d83d9c407964d159e1c72518c4d56935d779fc5d741a69304b
+size 497783424

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,14 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 1024,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}