Perfect7613/qwen3-8b-ih-grpo-tensorboard

Qwen3-8B IH-Challenge GRPO — Simple TensorBoard

This public dashboard is the plain-language view of the clean adaptive-attack Qwen3-8B runs:

• —GRPO-100-steps
• —GRPO-500-steps

Only 13 interpretable curves are shown. The most important group is 1. Main result:

• —Held-out accuracy (%): how often the model follows the correct instruction hierarchy on unseen examples.
• —Improvement over base (percentage points): held-out gain relative to step 0, before GRPO.
• —Training reward (%): reward on each sampled GRPO training batch; this is naturally noisy.

The other groups show accuracy by task, GRPO batch composition, and basic training-health checks.

For the 500-step run, held-out accuracy rises from 52.36% at step 0 to 81.89% at step 490, an improvement of 29.53 percentage points. The raw full-resolution logs remain in the model repository for research and audit use.

Model and dataset: https://huggingface.co/Perfect7613/qwen3-8b-ih-grpo-seed42