Spaces:
Paused
Paused
Commit History
-
Cleanup: drop legacy support-ops cruft now that DriftShield is the only collection
raj921After confirming the OpenEnv canonical layout (https://meta-pytorch.org/OpenEnv/guides/environment-anatomy.html) the legacy SaaS-support tasks were genuinely dead weight under the hard pivot to D1. This commit removes them. Deleted (-2333 lines): - tasks.py: 8 legacy TaskSpecs (c1*/c2*/c4*/c8*) and the COLLECTIONS entries for C1/C2/C4/C8. Removed legacy DIFFICULTY_TIERS dict. get_curriculum_task_ids() now only knows DriftShield aliases: driftshield, driftshield_easy, easy/medium/hard/expert (DriftShield ladder), all, or any task_id. - inference.py: 8 legacy fallback_action plans; only the 4 D1 plans remain. - tests/test_support_ops_env.py: entire file (every test was bound to c1_access_lockout / c4_gdpr_churn). - tests/test_driftshield.py: dropped the legacy comparison test that iterated over the 8 deleted task ids; rewrote the components-presence test to iterate DRIFTSHIELD_TASK_IDS instead of mixing collections. - __init__.py: stopped exporting DIFFICULTY_TIERS (legacy…3b79f14
-
feat: GRPO training stack (Qwen3 + Unsloth opt-in, curriculum, eval, audit)
raj921Adds the full RL training path for the Support Ops Env: * train.py: Qwen GRPO trainer with --difficulty curriculum, --use-unsloth opt-in path (unsloth/Qwen3.5-4B, bf16 LoRA), --load-in-4bit for T4, --use-vllm opt-in for A100. Default model Qwen/Qwen3-4B-Instruct-2507. * train_gemma4.py: experimental Gemma 4 path via TRL environment_factory. * tasks.py + __init__.py: curriculum helpers (easy/medium/hard/expert/all). * eval_compare.py: baseline-vs-trained evaluator -> JSON + Markdown table. * audit.py: reward-hack audit -> trajectory dumps + flag counts. * plot_rewards.py: 4-panel reward curves (total / components / penalty / rolling pass-rate). Reads v1 or v2 CSV schema. * pyproject.toml: [train] + [gemma] + [unsloth] optional-deps. * support_ops_colab.ipynb + support_ops_colab_gemma4.ipynb: full Colab notebooks with curriculum + Unsloth toggle + 4-bit fallback. * README.md: judge-facing rewrite aligned with the hackathon self-serve guide (tl;dr table, curriculum, reward desi…ffa6d1d
-
clean rewrite: shorter vars, no comments, same logic
raj921cbe3d95