Spaces:
Sleeping
Sleeping
Commit History
-
feat: environment redesign — real CSV data, shaped rewards, difficulty tiers
aatmk-panse- 65 products from 3 real Shopify CSV exports (apparel, apparel_2, jewelry) - 264 issue pool with random sampling per episode - Shaped rewards: discovery, partial fix, efficiency, regression, repetition - 3 hint tiers: full (easy), descriptions (medium), summary-only (hard) - Updated landing page and README329e3d3 -
fix: actionable hints now include real IDs (location_ids, product_ids, rules)
aatmk-panse- Inventory hints show exact location_id (loc_nyc) and quantity to set - Collection hints show which product_ids are missing - Rules hints show the exact corrected rule object - query_store_health returns available_location_ids and available_collection_ids2bcac83 -
fix: fallback to max_completion_tokens for newer models
aatmk-panseca40c95 -
fix: coerce string prices to float in update_variant
aatmk-panseLLMs send "49.99" (string) not 49.99 (number). The grader's isinstance check for (int, float) was failing silently. Now update_variant coerces price, compare_at_price, and weight to float before storing and grading.f824b2c -
fix: inference.py runs all 3 tasks with proper [START]/[STEP]/[END] per task
aatmk-panse- Loops over product_listing_qa, seo_collection_optimization, full_store_audit - Each task gets its own [START]...[END] log block with score in [0,1] - Supports IMAGE_NAME (Docker), ENV_URL (remote), or localhost:8000 (default) - Completes in <60s even without LLM (fallback actions)dce8bf6 -
feat: add /tasks and /grade endpoints for competition validator
aatmk-panse- GET /tasks — enumerates all 3 tasks with grader info (has_grader: true) - POST /grade — runs grader on a sample, returns score in [0, 1] - inference.py defaults to localhost:8000 (no Docker-in-Docker)c64fcea -
fix: move all files to repo root (HF Spaces requires README.md at root)
aatmk-panse699d744 -
Fix inference.py stdout format to match sample spec exactly
aatmk-panse- reward formatted as .2f, score as .3f - done/success as lowercase true/false (not Python True/False) - error as literal null (not Python None) - rewards as comma-separated .2f values (not Python list repr) - Align env var pattern with sample: IMAGE_NAME, API_KEY fallbackda3eda7 -
Add 3 tasks with graders and clamp scores to (0, 1)
aatmk-panse- Add tasks section to openenv.yaml with grader references for all 3 tasks (product_listing_qa, seo_collection_optimization, full_store_audit) - Create server/graders.py with grade functions and safe_reward clamping - Clamp health_score to [0.01, 0.99] in store.py, environment, and inference - Fixes "Not enough tasks with graders" and "scores out of range" validationc5968b0 -
Align env var names with OpenEnv sample inference spec
aatmk-panseUse os.getenv with exact variable names: API_BASE_URL, MODEL_NAME, HF_TOKEN, LOCAL_IMAGE_NAME matching the required convention.50cd687 -
Initial commit: Shopify store audit OpenEnv environment
aatmk-panse362bbff