GanitLLM-1.7B_SFT_GRPO

Highlights

GanitLLM-1.7B_SFT_GRPO is a compact Bengali mathematical reasoning model trained with SFT followed by standard GRPO. Key improvements over the base Qwen3-1.7B model:

+38.4 accuracy on Bn-MGSM benchmark (15.2 → 53.6)
+52.8 accuracy on Bn-MSVAMP benchmark (14.1 → 66.9)
88.32% Bengali reasoning (vs 19.64% for base model)
81.6% fewer tokens in generated solutions (1124 → 207 words)

Model Overview

Property	Value
Model Type	Causal Language Model
Base Model	Qwen/Qwen3-1.7B
Parameters	1.7B
Training	SFT + GRPO
Context Length	4,096 tokens
Language	Bengali, English

Training Details

This model was trained using a two-stage pipeline:

Supervised Fine-Tuning (SFT): Trained on GANIT-SFT (~11k examples) to ground reasoning in Bengali
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-1.7B_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-1.7B_SFT_GRPO --max-model-len 4096

Performance

Model	Bn-MGSM	Bn-MSVAMP	Avg. Words	Bengali %
Qwen3-1.7B (base)	15.20	14.10	1124	19.64%
GanitLLM-1.7B_SFT_GRPO	53.60	66.90	207	88.32%

Related Models

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