README.md

metadata

library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3-4B
pipeline_tag: text-generation
language:
  - bn
  - en
tags:
  - math
  - bengali
  - reasoning
  - grpo
datasets:
  - dipta007/Ganit

GanitLLM-4B_SFT_GRPO

Highlights

GanitLLM-4B_SFT_GRPO is a Bengali mathematical reasoning model trained with SFT followed by standard GRPO (without curriculum). Key improvements over the base Qwen3-4B model:

• +8.4 accuracy on Bn-MGSM benchmark (69.2 → 77.6)
• +5.8 accuracy on Bn-MSVAMP benchmark (70.5 → 76.3)
• 88.61% Bengali reasoning (vs 14.79% for base model)
• 80.0% fewer tokens in generated solutions (943 → 189 words)

Model Overview

Property	Value
Model Type	Causal Language Model
Base Model	Qwen/Qwen3-4B
Parameters	4B
Training	SFT + GRPO
Context Length	4,096 tokens
Language	Bengali, English

Training Details

This model was trained using a two-stage pipeline:

1. Supervised Fine-Tuning (SFT): Trained on GANIT-SFT (~11k examples) to ground reasoning in Bengali
2. GRPO: Standard reinforcement learning with random sampling on GANIT-RLVR (~7.3k examples)

Reward Functions

• Format Reward: Validates <think> and <answer> tag structure
• Correctness Reward: +2.0 for Bengali answer match, +1.0 for English match
• Bengali Reasoning Reward: Ensures >80% Bengali text in reasoning

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "dipta007/GanitLLM-4B_SFT_GRPO"

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_GRPO --max-model-len 4096

Performance

Model	Bn-MGSM	Bn-MSVAMP	Avg. Words	Bengali %
Qwen3-4B (base)	69.20	70.50	943	14.79%
GanitLLM-4B_SFT_GRPO	77.60	76.30	189	88.61%

Related Models

Model	Parameters	Training	Link
GanitLLM-4B_SFT_CGRPO	4B	SFT + CGRPO	Link
GanitLLM-4B_SFT_GRPO	4B	SFT + GRPO	Link
GanitLLM-4B_CGRPO	4B	CGRPO	Link
GanitLLM-1.7B_SFT_GRPO	1.7B	SFT + GRPO	Link
GanitLLM-0.6B_SFT_GRPO	0.6B	SFT + GRPO	Link

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.