README.md

metadata

title: RLM Model Evaluation
sdk: docker
hardware: t4-small

RLM Model Evaluation

Evaluates the trained needle-in-haystack model against the base model.

Models

• Base: Qwen/Qwen3-0.6B-Base
• Trained: mindchain/qwen3-0.6b-rlm-needle

Expected Results

• Base: ~25% accuracy (random guessing)
• Trained: 50-75% accuracy (after GRPO training)