Spaces:
Sleeping
Sleeping
Commit History
-
docs: add training.md operator quickstart, remove superpowers planning artifacts
Virajdocs/training.md mirrors inference.md's spirit: terse, command-first operator-facing quickstart for HF Jobs training runs. Covers smoke and real flavors, .secrets layout, empirical h200 cost table, the metrics that actually matter on a GRPO run (reward / reward_std / mean_length, not loss), and the env-client retry contract. docs/superpowers/{plans,specs} were one-shot planning artifacts from the initial training implementation; ~1.5k lines of historical scaffolding that add noise to the docs tree. The README and the new training.md cover what's actually still relevant.303e4f7
-
Add episode replay dashboard and GRPO training pipeline
Navaneeth SharmaReact/Vite/Tailwind dashboard with step-by-step episode playback, red/blue command feeds, metric charts, and service status display. Includes CSV-to-episode converter and Vite plugin for auto-generating scenario JSON from outputs/. Self-contained training/ package with env adapter, rollout collection, Unsloth + TRL GRPO trainer wiring, Hugging Face Job entrypoints, adapter and merged-model publishing, and Space deployment assets.509ef18 -
Update Red/Blue showdown behavior and refresh Qwen benchmark artifacts.
VirajEnforce one direct process-kill action per episode with explicit feedback, add clearer Blue runtime context, fix inference run documentation, and keep only the latest Qwen 3.5 9B output set. Made-with: Cursorf4ce885