Spaces:
Sleeping
Sleeping
Commit History
-
fix: always run all 3 task levels regardless of TASK_LEVEL env var
ShivappaRoot cause: Dockerfile had ENV TASK_LEVEL=easy which caused inference.py to only run the easy task — validator saw 1 task, needs 3. - Dockerfile: remove ENV TASK_LEVEL=easy entirely - inference.py: always run [easy, medium, hard] unconditionally; TASK_LEVEL env var only affects the uvicorn server, not inference82f2af5
-
fix: run ALL THREE task levels (easy/medium/hard) in inference.py
ShivappaThe validator requires at least 3 tasks with graders. Previously inference.py only ran a single TASK_LEVEL (defaulting to 'easy'), so the validator saw only 1 task — failing 'Not enough tasks with graders'. Changes: - Extract run_episode() helper to run one task level cleanly - main() now runs easy + medium + hard sequentially by default - Each task gets its own [START]..[STEP]..[END] output block - If TASK_LEVEL env var is set to a specific level, only that level runs (preserves local testing behaviour)f15c8b1 -
fix: clamp all scores strictly to (0.01, 0.99) — validator rejects 0.0 and 1.0
Shivappagraders.py: - Add _clamp() helper: min(max(score, 0.01), 0.99) - All three graders (easy/medium/hard) now return _clamp(score) - Scores can never be exactly 0.0 or 1.0 my_environment.py: - Clamp per-step reward to (0.01, 0.99) after penalty calculation - Reset observation uses 0.01 instead of 0.0 inference.py: - Final episode score clamped to (0.01, 0.99) before logging [END] line1816098 -
fix: remove silent except in get_model_action + fix port default
Shivappa- Remove try/except in get_model_action() — LLM errors now propagate loudly so the validator log shows the real failure reason - Fix ENV_BASE_URL default: 7860 (matches Dockerfile PORT) not 8000 - No dotenv, no fallback model, no silent swallowing of API errorscaa07e1 -
fix: guarantee LLM calls go through validator proxy in inference.py
ShivappaCritical changes: - Remove load_dotenv() entirely — no .env file can interfere with injected vars - Move API_KEY + API_BASE_URL reads inside main() using os.environ.get() so they are resolved at runtime, not at import time - API_BASE_URL has NO hardcoded default — if validator injects it, it is used; fallback only applies for local testing with explicit warning printed - Change default MODEL_NAME from Qwen/Qwen2.5-72B-Instruct to gpt-4o-mini (LiteLLM proxy supports OpenAI model names, not HF router paths) - Print API_BASE_URL and MODEL_NAME at startup for validator log visibility - Improve LLM error logging to show exception type and message6026955 -
update env
ef42e68 -
fix: copy inference.py into Docker image and fix API_KEY precedence
Shivappa- Dockerfile: copy inference.py into image + install openai & python-dotenv - inference.py: load_dotenv(override=False) so injected API_KEY/API_BASE_URL are never overwritten by a local .env file - inference.py: fix lookup order — API_KEY checked before HF_TOKEN - inference.py: fix error message to reference API_KEY not HF_TOKEN Fixes 'No API calls were made through LLM proxy' validator error.ed7afa9 -
fix: use API_KEY and API_BASE_URL env vars injected by hackathon validator
Shivappa- inference.py: prioritize API_KEY over HF_TOKEN in lookup order - baseline.py: replace OPENAI_API_KEY/OPENAI_BASE_URL with API_KEY/API_BASE_URL - both scripts now route all LLM calls through the validator's LiteLLM proxy - fallback chain: API_KEY -> HF_TOKEN -> OPENAI_API_KEY for local dev compatd2b1c92 -
Email triage RL
753b5ee