You can deploy this model and start earning money today!

friendshipkim/
Qwen2.5-Math-1.5B-Scoring-Mean

Name: Qwen2.5-Math-1.5B-Scoring-Mean
Availability: InStock

safetensors

qwen2

reward-model

success-rate-prediction

custom

custom_code

friendshipkim/Qwen2.5-Math-1.5B

finetune

apache-2.0

Views

No views yet

Qwen2.5-Math-1.5B-Scoring-Mean

This is a custom Qwen2 model with dual heads:

Language Model Head: Standard next-token prediction for text generation
Success Rate Head: Predicts a success probability score in [0, 1] for the sequence

Base Model

This model is based on friendshipkim/Qwen2.5-Math-1.5B.

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model with trust_remote_code=True
model = AutoModelForCausalLM.from_pretrained(
    "friendshipkim/Qwen2.5-Math-1.5B-Scoring-Mean",
    trust_remote_code=True,
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("friendshipkim/Qwen2.5-Math-1.5B-Scoring-Mean")

# Example: Get both LM output and success score
prompt = "Question: What is 2+2?\nAnswer: 4"
inputs = tokenizer(prompt, return_tensors="pt")

# Get both outputs
lm_output, success_score = model(**inputs, return_score=True)
print(f"Success rate: {success_score.item():.3f}")

# Generate text (return_score=False for standard generation)
generated = model.generate(**inputs, max_length=50, return_score=False)
print(tokenizer.decode(generated[0]))

Model Architecture

Backbone: Qwen2 transformer model
LM Head: Linear layer for next-token prediction (vocab_size outputs)
Success Rate Head: Linear layer for sequence scoring (1 output, sigmoid activation)

Training

The success_rate_head is randomly initialized and needs to be fine-tuned on your task. The LM head and backbone are initialized from the base model.

Custom Modeling

This model uses a custom modeling file (modeling_custom.py) that extends Qwen2ForCausalLM. The return_score parameter controls whether to compute the success rate:

return_score=True: Returns (lm_output, success_score)
return_score=False: Returns lm_output only (for standard generation)