GanitLLM-4B-SFT

GanitLLM-4B_SFT

<p align="center"> <a href="https://arxiv.org/abs/2601.06767"> <img src="https://img.shields.io/badge/%F0%9F%94%A5_Accepted_at-ACL_2026_(Findings)_%F0%9F%94%A5-b12a00?style=for-the-badge&labelColor=ffb300" alt="Accepted at ACL 2026 (Findings)"> </a> </p>

Highlights

GanitLLM-4B_SFT is a Bengali mathematical reasoning model trained with Supervised Fine-Tuning on the GANIT dataset. This model serves as the foundation for further RL training (GRPO/CGRPO). Key improvements over the base Qwen3-4B model:

• +4.80 accuracy on Bn-MGSM benchmark (69.20 → 74.00)
• +4.10 accuracy on Bn-MSVAMP benchmark (70.50 → 74.60)
• 86.65% Bengali reasoning (vs 14.79% for base model)
• 80.5% fewer words in generated solutions (943 → 184 words)

Note: This is the SFT-only checkpoint. For best results, use the RL-enhanced versions: GanitLLM-4B_SFT_CGRPO or GanitLLM-4B_SFT_GRPO.

Model Overview

Training Details

This model was trained with a single-stage pipeline:

1. Supervised Fine-Tuning (SFT): Trained on GANIT-SFT (~11k examples) to ground reasoning in Bengali

Training Data

• Dataset: GANIT-SFT (11,023 examples)
• Format: Bengali math problems with chain-of-thought reasoning
• Structure: <think> tags for reasoning, <answer> tags for final answer

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "dipta007/GanitLLM-4B_SFT"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

problem = "একটি দোকানে ১২টি আপেল আছে। যদি ৫টি আপেল বিক্রি হয়, তাহলে কতটি আপেল বাকি থাকবে?"

prompt = f"""A conversation takes place between the user and the assistant. The user asks a question, and the assistant solves the problem. Please reason step by step in Bengali, and put your final answer in the <answer> </answer> tags.

Question: {problem}"""

messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(**model_inputs, max_new_tokens=2048, temperature=0.7)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
response = tokenizer.decode(output_ids, skip_special_tokens=True)
print(response)

Using vLLM

vllm serve dipta007/GanitLLM-4B_SFT --max-model-len 4096

Performance

Related Models

Citation

@inproceedings{dipta2026ganitllm,
  title={GanitLLM: Difficulty-Aware Bengali Mathematical Reasoning through Curriculum-GRPO},
  author={Shubhashis Roy Dipta and Khairul Mahbub and Nadia Najjar},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2026},
  year={2026},
  eprint={2601.06767},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2601.06767},
}

License

This model is released under the Apache 2.0 License.