README.md

---
base_model:
- Qwen/Qwen2.5-3B-Instruct
pipeline_tag: text-generation
library_name: transformers
---
# AbleCredit Reasoner R0 Qwen 2.5 3B Instruct

## Introduction

This model is trained by Deepseek R1 style (GRPO) reinforcement learning on Qwen 2.5 3B Instruct as a base model.
Primarily intended for research in application of small LLMs trained using GRPO/RL in the domain of finance, credit underwriting etc.

### Model Description

- **Fine Tuned by:** AbleCredit (LightBees Technologies Private Limited, Bengaluru, India)
- **License:** We've retained the original Qwen research license. Note that license does not allow commercial use.
- **Finetuned from model:** Qwen/Qwen2.5-3B-Instruct

## How to Get Started with the Model

Use with standard Huggingface based setup

```python
model_name = "AbleCredit/AbleCredit-R0-Qwen-2.5-3B-Instruct" # or local path to model
system_prompt = {
    "role": "system",
    "content": (
        "You are a helpful assistant. User asks a question the assistant answers it.\n"
        "The assistant first thinks about reasoning process in mind and then provides the user with the answer."
        ),
      }

suffix_prompt = {
    "role": "assistant",
    "content": "Let me solve this step by step.\n<think>",
}

prompt_msgs = [
    system_prompt,
    {"role": "user", "content": "What is 15 times 3 ?"},
    suffix_prompt,
]

base_model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = tokenizer.apply_chat_template(
    prompt_msgs,
    tokenize=False,
    continue_final_message=True,
    add_generation_prompt=False,
)

# Tokenize the prompt and move it to the appropriate device.
inputs = tokenizer(prompt, return_tensors="pt").to("cuda:0")

print("\nGenerating response...\n")
outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.5,
    min_p=0.01,
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print("\nResponse:\n", response)
```

## Training Details

### Training Data

Trained using open source logical reasoning datasets and a proprietary finance dataset created by AbleCredit.com.

### Training Procedure

Trained using deepseek style reinforcement learning using GRPO with rule based rewards.

## Evaluation

- Model achieves ~67% score on GSM8K benchmark in a **zero shot** setting (check benchmarking script for more details).

## Model Card Contact

[contact Harshad Saykhedkar via LinkedIn](https://www.linkedin.com/in/harshadss/)