Spaces:
Running
Running
Commit History
-
Add real GRPO eval data: per-step reward chart from 42 episodes
Replaced placeholder RL chart with actual data from grpo_eval results. Shows avg and max reward per episode step (steps 1-14). Step 1 gets +0.72 avg (exploration bonus), diminishing to near-zero by step 8. 42 episodes, 0 solved, but non-zero gradient signal throughout. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1717221 -
Initial release: NetOps RL environment for network operations
Environment: 37 fault types, 4 difficulty levels, 7-component shaped reward, pure-Python simulator (192x faster than Mininet, no root required). Frontend: Gradio app with interactive topology, live simulator integration, guided walkthroughs, leaderboard (13 models), training logs, fleet vision page. Training: 371 expert episodes from Sonnet distillation, SFT on Qwen3-1.7B (loss 0.84→0.01), GRPO RL pipeline. Research grounding: Gill SIGCOMM 2011, Meza IMC 2018, Goldberg CACM 2014. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1195f0c