erayalp/qwen2.5-0.5b-instruct-GRPO-v3-tr-math-gsm8k

Objective

This model is the final product of the multi-stage training pipeline designed to improve the Turkish mathematical reasoning capabilities of the compact Qwen2.5-0.5B model.

Starting from erayalp/qwen2.5-0.5b-instruct-sft-v2-tr-math-medium, which was fine-tuned on 2-3 step reasoning Turkish math problems, this version continues training using ytu-ce-cosmos/gsm8k_tr to improve the model’s step-by-step reasoning and generalization to multi step grade school math performance on such small model.

This model is intended for:

• Researchers exploring reinforcement learning on small LLMs.
• Research on curriculum learning and multi-step math reasoning in small models.
• Comparative baselines for evaluating Turkish math reasoning tasks of grade school math complexity.

Limitations

• With only 0.5B parameters, it may not perform as robustly as larger models.
• Math-specific hallucinations may persist in underrepresented edge patterns.
• Prompt sensitivity and reasoning depth are open to future improvements.

Roadmap

1. Phase 1: SFT with basic arithmatic and math problems
2. Phase 2: SFT with moderately difficult math problems
3. Phase 3: SFT with full-scale GSM8K-TR complexity
4. 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:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "erayalp/qwen2.5-0.5B-instruct-GRPO-v3-tr-math-gsm8k"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Bir bahçede 25 gül var. 40 lale vardır. 35 papatya var. Çiçeklerin yüzde kaçı gül değildir?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))