Instructions to use jihwan1205/svp-v-coconut-gpt2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jihwan1205/svp-v-coconut-gpt2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="jihwan1205/svp-v-coconut-gpt2")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("jihwan1205/svp-v-coconut-gpt2")
model = AutoModelForCausalLM.from_pretrained("jihwan1205/svp-v-coconut-gpt2", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jihwan1205/svp-v-coconut-gpt2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jihwan1205/svp-v-coconut-gpt2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jihwan1205/svp-v-coconut-gpt2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/jihwan1205/svp-v-coconut-gpt2

SGLang

How to use jihwan1205/svp-v-coconut-gpt2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "jihwan1205/svp-v-coconut-gpt2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jihwan1205/svp-v-coconut-gpt2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "jihwan1205/svp-v-coconut-gpt2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jihwan1205/svp-v-coconut-gpt2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use jihwan1205/svp-v-coconut-gpt2 with Docker Model Runner:
```
docker model run hf.co/jihwan1205/svp-v-coconut-gpt2
```

jihwan1205 commited on 3 days ago

Commit

7554e05

verified ·

1 Parent(s): be4cc8d

Align model card and weights with ICLR 2027 paper

Browse files

Files changed (2) hide show

README.md +12 -6
model.safetensors +1 -1

README.md CHANGED Viewed

@@ -19,13 +19,19 @@ datasets:
 A [COCONUT](https://huggingface.co/ModalityDance/latent-tts-coconut) GPT-2 (124M)
 latent-reasoning model post-trained with **SVP-V-GRPO** — reinforcement learning whose
 rollout exploration comes from *weight-space* perturbation of the attention value
-projections (σᵢ → σᵢ(1+αgᵢ) on the singular values of `W_V`) rather than from token
-sampling.
 Only the value-projection weights (the V columns of each `c_attn`) differ from the base
 checkpoint. The perturbation is a training-time exploration mechanism and is **not** used
 at deployment: inference is ordinary greedy decoding.
 ## Results
 Clean greedy decoding, `max_new_tokens=64`, exact-match on the final answer; all rows
@@ -69,11 +75,11 @@ COCONUT inference code works as well.
 | | |
 |---|---|
 | Base | `ModalityDance/latent-tts-coconut` (COCONUT GPT-2, 124M) |
-| Data | [GSM8K-Aug](https://huggingface.co/datasets/zen-E/GSM8k-Aug) training split (385,620 problems) |
-| Algorithm | GRPO — G=32 rollouts/prompt, 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` (α=0.6), held fixed across prefill and all latent steps, off at answer decoding |
 | Trained parameters | V columns of every `c_attn` (≈ 7M of 124M) |
-| Optimizer | AdamW, constant lr, grad-clip 1.0 |
 | Reward | Final-answer correctness only |
 ## Limitations

 A [COCONUT](https://huggingface.co/ModalityDance/latent-tts-coconut) GPT-2 (124M)
 latent-reasoning model post-trained with **SVP-V-GRPO** — reinforcement learning whose
 rollout exploration comes from *weight-space* perturbation of the attention value
+projections. For each layer and rollout, SVP independently samples Gaussian noise and
+rescales coefficients in the fixed SVD basis of `W_V` as `sᵢ → sᵢ(1+αgᵢ)`. The resulting
+coefficients can be negative and are not necessarily singular values of the perturbed
+matrix.
 Only the value-projection weights (the V columns of each `c_attn`) differ from the base
 checkpoint. The perturbation is a training-time exploration mechanism and is **not** used
 at deployment: inference is ordinary greedy decoding.
+This repository contains the paper's GPT-2 SVP-V-GRPO checkpoint from epoch 9 of the
+15-epoch `B=32`, `lr=6e-5`, seed-0 run. The results below are the paper's Figure 3
+evaluations; they have not been independently rerun for this model-card update.
 ## Results
 Clean greedy decoding, `max_new_tokens=64`, exact-match on the final answer; all rows
 | | |
 |---|---|
 | Base | `ModalityDance/latent-tts-coconut` (COCONUT GPT-2, 124M) |
+| Data | 383,891 eligible prompts from [GSM8K-Aug](https://huggingface.co/datasets/zen-E/GSM8k-Aug) |
+| Algorithm | GRPO: `G=32` rollouts per prompt; `2B=64` candidate prompts per iteration with up to `B=32` mixed-outcome groups retained; `μ=2` optimization passes; centered advantages; symmetric clipping `ε=0.2`; k3 KL (`β=0.02`) to the frozen base |
+| Exploration | SVP with `α=0.6`: independent Gaussian coefficient perturbations for each Value-projection layer and rollout, held fixed through prefill and all six latent steps; clean answer decoding; SVD factors refreshed every 50 iterations |
 | Trained parameters | V columns of every `c_attn` (≈ 7M of 124M) |
+| Optimizer | AdamW, constant `lr=6e-5`, seed 0, no warmup or weight decay, gradient clipping at 1.0; epoch-9 checkpoint from 15 epochs |
 | Reward | Final-answer correctness only |
 ## Limitations

model.safetensors CHANGED Viewed

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

 version https://git-lfs.github.com/spec/v1
+oid sha256:0b1d2bd8d54ca467cf05339562173a4e8f0742c4a84e24059cfdef6b8ea05402
 size 497783424