README.md

---
library_name: transformers
pipeline_tag: text-generation
tags:
- value-guided-search
---

# Model Card for Model ID

1.5B value model for guiding DeepSeek CoT: arxiv.org/abs/2505.17373.

[Value-Guided Search for Efficient Chain-of-Thought Reasoning](https://huggingface.co/papers/2505.17373)

Code: https://github.com/kaiwenw/value-guided-search

This model is a `Qwen2ForClassifier` model, a modified version of the Qwen2 model for classification tasks, which is used to guide chain-of-thought reasoning.

## Usage

To load the model, you can use the following code snippet:

```python
import classifier_lib
import torch

model_loading_kwargs = dict(attn_implementation="flash_attention_2", torch_dtype=torch.bfloat16, use_cache=False)
classifier = classifier_lib.Qwen2ForClassifier.from_pretrained("VGS-AI/DeepSeek-VM-1.5B", **model_loading_kwargs)
```

To apply the model to `input_ids`, you can use the following code snippet:

```python
import torch

device = torch.device("cuda")
# your input_ids
input_ids = torch.tensor([151646, 151644, 18, 13, 47238, ...], dtype=torch.long, device=device)
attention_mask = torch.ones_like(input_ids)
classifier_outputs = classifier(input_ids.unsqueeze(0), attention_mask=attention_mask.unsqueeze(0))
# use last index of the sequence
scores = classifier_outputs.success_probs.squeeze(0)[-1].item()
```