|
--- |
|
library_name: transformers |
|
tags: |
|
- reward-model |
|
- prm |
|
- generative reward model |
|
- process supervision |
|
- chain-of-thought |
|
- verification |
|
- math reasoning |
|
- code verification |
|
license: apache-2.0 |
|
pipeline_tag: text-generation |
|
--- |
|
|
|
# Model Card for ThinkPRM-7B |
|
|
|
ThinkPRM-7B is a generative Process Reward Model (PRM) based on the R1-Distill-Qwen-7B architecture. It is fine-tuned to perform step-by-step verification of reasoning processes (like mathematical solutions) by generating an explicit verification chain-of-thought (CoT) that involves labeling every step. It is designed to be highly data-efficient, requiring significantly less supervision data than traditional discriminative PRMs while achieving strong performance. |
|
|
|
Here's an example of the model output: |
|
|
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
ThinkPRM-7B provides step-level verification scores by generating natural language critiques and correctness judgments for each step in a given solution prefix. It leverages the underlying reasoning capabilities of the base Large Reasoning Model (LRM) and enhances them through fine-tuning on a small (1K examples) dataset of synthetically generated verification CoTs. These synthetic CoTs were produced by prompting QwQ-32B-Preview and filtered against ground-truth step labels from the PRM800K dataset to ensure quality. |
|
|
|
The model uses a standard language modeling objective, making it interpretable and allowing it to scale process verification compute by generating longer or multiple verification CoTs. It demonstrated superior performance compared to LLM-as-a-judge and discriminative PRM baselines (based on the same R1-Distill-Qwen-7B model but trained on ~100x more labels) on benchmarks including ProcessBench, MATH-500, AIME '24, GPQA-Diamond, and LiveCodeBench. |
|
|
|
- **Finetuned from model [optional]:** [R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) |
|
|
|
### Model Sources [optional] |
|
|
|
- **Repository:** [Github](https://github.com/mukhal/thinkprm) |
|
- **Paper:** [Process Reward Models that Think (arXiv:2504.16828)](https://arxiv.org/abs/2504.16828) |
|
|
|
|
|
### Direct Use |
|
|
|
ThinkPRM-7B is intended for verifying the correctness of step-by-step reasoning processes. Primary uses include: |
|
- **Scoring Solutions:** Assigning step-level or overall scores to candidate solutions for ranking in Best-of-N sampling or guiding tree search in reasoning tasks. |
|
- **Generating Verification Rationales/CoTs:** Producing detailed chain-of-thought verifications that explain *why* a particular step is correct or incorrect, aiding interpretability. |
|
- **Standalone Verification:** Evaluating the correctness of a given problem-solution pair. |
|
|
|
The model has been evaluated on mathematical reasoning (MATH, AIME), scientific QA (GPQA), and code generation (LiveCodeBench). See our paper for more details. |
|
|
|
## Limitations |
|
|
|
- **Overconfidence:** Generative PRMs like ThinkPRM can sometimes produce scores clustered near 0 or 1, potentially not reflecting true uncertainty |
|
- **Step Label Interference:** The autoregressive nature might cause an early incorrect step judgment to negatively bias the evaluation of subsequent steps. |
|
- **Sensitivity to Formatting/Prompting:** Performance might be sensitive to the exact format of the input solution and the prompt used for verification (though fine-tuning likely reduces this compared to LLM-as-a-judge). |
|
|
|
## How to Get Started with the Model |
|
```python |
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
from vllm import LLM, SamplingParams |
|
|
|
model_id = "launch/ThinkPRM-7B" # Replace with actual model ID on Hub |
|
tokenizer = AutoTokenizer.from_pretrained(model_id) |
|
llm = LLM(model=model_id, max_model_len=16384) |
|
|
|
# Example problem and solution |
|
problem = "Solve for x: 2x + 3 = 7" |
|
prefix = "Step 1: Subtract 3 from both sides: 2x = 4 |
|
Step 2: Divide by 2: x = 1" |
|
|
|
# Format the prompt |
|
prompt = f"""You are given a math problem and a proposed step-by-step solution: |
|
|
|
[Math Problem] |
|
|
|
{problem} |
|
|
|
[Solution] |
|
|
|
{prefix} |
|
|
|
Review and critique each step in the proposed solution to determine whether each step is correct. If the solution is incomplete, only verify the provided steps |
|
""" |
|
|
|
prompt = tokenizer.apply_chat_template([ |
|
{'role': "user", "content": prompt} |
|
], tokenize=False, add_generation_prompt=True) + " |
|
Let's verify step by step:" |
|
|
|
# Set sampling parameters |
|
sampling_params = SamplingParams( |
|
temperature=0.0, |
|
max_tokens=4096, |
|
stop=None |
|
) |
|
|
|
# Generate the verification |
|
outputs = llm.generate(prompt, sampling_params) |
|
verification_cot = outputs[0].outputs[0].text |
|
|
|
print(verification_cot) |
|
""" |
|
Step 1: Subtract 3 from both sides: 2x = 4 |
|
|
|
Critique: Starting with the equation 2x + 3 = 7, subtracting 3 from both sides is a correct operation to isolate the term with the variable. So, 2x + 3 - 3 = 7 - 3, which simplifies |
|
to 2x = 4. This step seems correct. |
|
|
|
Step 2: Divide by 2: x = 1 |
|
|
|
Critique: Now, to solve for x, we need to divide both sides of the equation by 2. So, 2x / 2 = 4 / 2, which simplifies to x = 2. Wait a minute, the solution says x = 1, but accordin |
|
g to this calculation, it should be x = 2. This seems incorrect. |
|
|
|
Therefore, the first step is correct, but the second step has an error. |
|
|
|
**Final Output:** |
|
|
|
Let's verify step by step: |
|
|
|
Step 1: Subtract 3 from both sides: 2x = 4 |
|
|
|
Critique: This step is correct. Subtracting 3 from both sides of the equation 2x + 3 = 7 properly isolates the term with the variable, resulting in 2x = 4. |
|
|
|
Step 1 is \boxed{correct} |
|
|
|
Step 2: Divide by 2: x = 1 |
|
|
|
Critique: This step is incorrect. Dividing both sides of the equation 2x = 4 by 2 should yield x = 2, not x = 1. |
|
|
|
Step 2 is \boxed{incorrect} |
|
</think> |
|
Is the solution correct? No |
|
""" |