# Value Model (Base + Value Head)

- Base: Qwen/Qwen2.5-Math-1.5B-Instruct
- This folder contains base model weights (safetensors shards) and an extra `value_head.safetensors`.

## Quick inference (Python)

```python
import os
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from safetensors.torch import load_file

model_dir = "/home/huimin/New-Proj/value_model-1.5b/hf_converted_model"
base = AutoModelForCausalLM.from_pretrained(model_dir, torch_dtype=torch.bfloat16)
value_head = torch.nn.Linear(base.config.hidden_size, 1, bias=False)
state = load_file(os.path.join(model_dir, "value_head.safetensors"))
value_head.load_state_dict({"weight": state["value_head.weight"]})

tok = AutoTokenizer.from_pretrained(model_dir)
inputs = tok("Hello", return_tensors="pt")
outputs = base(**inputs, output_hidden_states=True)
last = outputs.hidden_states[-1]
values = value_head(last).squeeze(-1)
print(values.shape)
```