blackhatbadshah/scaler-hackathon-meta-hf

ai zigzag issue fix

fixed faliour collapse

Added context aware reward shaping

recalibrate score to steps taken instead of max steps

fixed reward score step logging

reward bounded to -5 and 5

fixed runtime config

Refactor and fixes

Readme fix

final cleanup

fixed wait action

loop handled vs unhandled emails more robustly in case of out-of-order processing

loop removed from inference.py, and updated runtime_config.py to check for both HF and OpenAI API keys. Also updated the default model in config.py to a Qwen model.

wait action removed

removed hurestic agent for final submition

scoring logic fix

fix openenv dependency

LLM OBSERVATION FILTER ADDED

loop fix

hf fix

openapiintegration added

hf router fixes

FIxed missing links

prompt update

step inddex Update

model fix

Response format fix

FIxed response format

fix Classifier logic

Correct runtime and inference documentation

Reflect inference score in reward quality

Align benchmark runtime with submission contract

Simplify LLM training and grading flow

Clean benchmark entrypoints and stale code paths

Surface model failure reasons in inference logs

Support runtime env overrides consistently

Prefer OpenAI backend when runtime config is present

Centralize runtime configuration access

Refine inference compatibility and logging defaults

Apply OpenEnv compliance fixes and expand sample tasks

run inference across canonical tasks

fix inference classify-once loop

align inference stdout contract

adjust inference backend fallback

refine benchmark submission surface

feat(tasks): updated agent to be more deterministic

feat(tasks): add validated json scenario loader

feat(rl): add trainer-facing adapter layer

feat(tasks): expand inbox cases

feat(tasks): expand sample registry