The goal of this project is to enhance the reasoning ability of the compact Qwen2.5-0.5B model on Turkish math questions. Using supervised fine-tuning (SFT) on simpler examples as a starting point, the model will be progressively improved through curriculum learning, and later refined using Group Relative Policy Optimization (GRPO) to boost multi-step reasoning performance.

This model is intended for:

Research on curriculum learning in small models
Evaluating Turkish math reasoning tasks

Limitations

Currently only trained on simpler math examples — lacks robustness for multi-step or abstract reasoning.
May produce incorrect or overconfident answers on complex tasks.
Performance may be sensitive to prompt phrasing.

Roadmap

Phase 1: SFT with basic arithmatic and math problems
Phase 2: SFT with moderately difficult math problems
Phase 3: SFT with full-scale GSM8K-TR complexity
Phase 4: GRPO-based training to optimize multi-step reasoning and reduce hallucinations

How to Use

You can easily run inference using the Transformers library:

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "erayalp/qwen2.5-0.5b-instruct-sft-v1-tr-math-easy"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Ali’nin 3 kalemi vardı. 2 kalem daha aldı. Ali’nin şimdi kaç kalemi var?"

inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=256)

print(tokenizer.decode(output[0], skip_special_tokens=True))

erayalp/qwen2.5-0.5b-instruct-SFT-v1-tr-math-easy

Objective

This model is intended for:

Limitations

Roadmap

How to Use

erayalp/
qwen2.5-0.5b-instruct-SFT-v1-tr-math-easy