jaydeepshah2025/data-quality-env

Fix 3 issues from organizer feedback: add score= to END, use API_KEY env var, remove API_BASE_URL default

Fix: clamp cumulative score too - sum of step rewards was hitting exactly 1.0

Ensure step reward is strictly (0,1): min 0.01, max 0.99 - never 0.0 or 1.0

Revert to standard OpenEnv reward format (raw values, null on reset) matching reference projects

Rewrite inference to support LOCAL_IMAGE_NAME (Docker) + fallback to HF Space URL

Add debug prints for ENV_URL and MODEL_NAME

Fix: default ENV_URL to live HF Space instead of localhost - validator has no local server

Add Dockerfile to server/ (OpenEnv standard), make PORT configurable via env var

Fix: return per-step reward (not cumulative) in reward field - prevents sum(rewards) > 1

Prevent cumulative score overflow (cap at 0.95) + remove negative rewards

Return cumulative score as reward field (not step reward) - ensures validator always sees clamped value

Fix environment: MIN_SCORE=0.02, MAX_SCORE=0.98, init total_reward=0.02, round to 2dp

Switch to key=value log format, remove score from END, remove json.dumps from logs

Final fix: normalize sum(rewards) to max 0.90, safe_reward (0.02-0.98), prevent all edge cases

Revert to JSON log format, _clamp every numeric value in all log lines

Rewrite inference.py: fix log format to key=value (was JSON), clamp all rewards

Use environment cumulative_score as task score instead of sum(rewards)

Use 1 worker to fix session routing - multi-worker breaks HTTP session state

Nuclear fix: clamp ALL scores/rewards everywhere with _clamp() helper

Move score clamping OUTSIDE try block - fixes score=0.0 on any error

Clamp every step reward to (0.001, 0.99) and cumulative to (0.001, 0.99) - fix negative rewards from false positives

Add base reward (0.001) to every step, scale by 0.95 - ensures all rewards in (0,1)

Scale rewards to 0.98x + base 0.005 so sum is always in (0, 1)

Fix: return step_reward (not cumulative) in reward field, clamp total internally

Clamp all environment scores to (0.01, 0.99) - validator requires open interval

Use AWS ECR public mirror to bypass Docker Hub registry failures

Try python:3.11-slim-bookworm for Docker Hub registry compatibility

Switch to python:3.12-slim to workaround Docker Hub registry errors

Clamp task scores to (0, 1) open interval per validator requirement

Pin Python image to 3.11.12-slim to fix Docker registry manifest error

Fix openenv validate: add uv.lock, server entry point, openenv-core dep

Add 'What Makes This Different' section to README

Remove default for HF_TOKEN in inference.py

Enhance README with stronger positioning and client examples

Restructure to standard OpenEnv layout

Initial commit: DataQual OpenEnv environment