chris0809/tinyLLM-0.51B-ARC-GRPO

tinyLLM adapter: arc

中文项目文档 · English project overview · Live demo

Base model: chris0809/tinyLLM-0.51B-SFT

This native tinyLLM LoRA is the best ARC-Easy GRPO checkpoint (step 200). Full-set accuracy improved from 30.09% to 35.98% (+5.89 percentage points).

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("chris0809/tinyLLM-0.51B-SFT", trust_remote_code=True)
model.load_lora_pretrained("chris0809/tinyLLM-0.51B-ARC-GRPO")

For ARC-style inference, render the question with the base tokenizer using enable_thinking=True. The template then injects tinyLLM's training-time reasoning protocol automatically. Append Put your final answer in LaTeX boxed form like $\boxed{answer}$. when reproducing the project's boxed-answer evaluation.

The adapter uses tinyLLM's native multi-adapter implementation and is stored with safetensors; it is not a PEFT adapter.