qwen05b-gsm8k-sft-instruct/README.md

license, base_model, datasets, language, pipeline_tag, library_name, tags

license	base_model	datasets	language	pipeline_tag	library_name	tags
apache-2.0	Qwen/Qwen2.5-0.5B-Instruct	openai/gsm8k	en	text-generation	transformers	gsm8k math sft qwen2.5

Qwen2.5-0.5B-Instruct — GSM8k SFT

Full supervised fine-tune of Qwen/Qwen2.5-0.5B-Instruct on the GSM8k training split (7,473 worked solutions), trained with verl's sft_trainer.

Results

Metric	Value
GSM8k test accuracy (greedy, 1319 examples)	36.2% (477/1319)

Training

• Base model: Qwen/Qwen2.5-0.5B-Instruct
• Method: full fine-tune (no LoRA), bf16 mixed precision
• Data: GSM8k train, chat format (system + user); assistant target is the full worked solution ending in #### N
• Epochs: 3 (348 steps); AdamW, lr 1e-5 cosine, warmup ratio 0.1, weight decay 0.1, grad clip 1.0
• Max sequence length: 1024

Prompt format

System prompt used during training; the model ends its answer with #### <number>:

You are a math problem solver. Think step by step. You MUST end your response with '#### ' where is the final numerical answer (digits only, no units or markdown).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

m = AutoModelForCausalLM.from_pretrained("OhhMoo/qwen05b-gsm8k-sft-instruct")
tok = AutoTokenizer.from_pretrained("OhhMoo/qwen05b-gsm8k-sft-instruct")

messages = [
    {"role": "system", "content": "You are a math problem solver. Think step by step. You MUST end your response with '#### <number>' where <number> is the final numerical answer (digits only, no units or markdown)."},
    {"role": "user", "content": "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May? Let's think step by step."},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
out = m.generate(ids, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))