You can deploy this model and start earning money today!

woojin93/
qwen0_5b-gsm8k-20251101-1022

Name: qwen0_5b-gsm8k-20251101-1022
Availability: InStock

safetensors

qwen2

Views

No views yet

Qwen2.5-0.5B — LoRA for GSM8K (Colab/T4-friendly)

Reproducible small-model math reasoning fine-tuning with LoRA.

Base: Qwen/Qwen2.5-0.5B
Training: 1 epoch, bs=1, grad_acc=8, lr=1e-4, warmup=0.03
Max length: 896 (example)
Eval: EleutherAI lm-eval-harness (GSM8K, fewshot=5, limit=50, hf backend)

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("woojin93/qwen0_5b-gsm8k-20251101-1022")
m = AutoModelForCausalLM.from_pretrained("woojin93/qwen0_5b-gsm8k-20251101-1022", torch_dtype="auto", device_map="auto")

prompt = "<|im_start|>user\nSolve: 12 + 35 = ?\n<|im_end|>\n<|im_start|>assistant\n"
out = m.generate(**tok(prompt, return_tensors='pt').to(m.device), max_new_tokens=64)
print(tok.decode(out[0], skip_special_tokens=True))