Spaces:
Sleeping
Sleeping
Commit History
-
Add real-source ingestion and expand hybrid corpus
c07553d -
Simplify provider config and improve UI guidance
fc6d479 -
Inline Space frontend into app
4f1b3fa -
Improve Space UI and compact END logs
f12561b
-
Improve Hugging Face UI layout
5819e51 -
Align runtime deps with HF Gradio extras
4c42e45 -
Quiet Gradio 6 app config
2366e5e -
Remove unused heavy runtime deps
5e5daea -
Align Gradio deps with HF builder
d7b03af -
Protect .env from agents
a061f2a -
Build Gradio Space frontend
fd2d750
-
Harden reward boundary handling
aa0019f -
Sync local Avigilance workspace
e1e3120 -
Fix frontend accessibility label and Space frontend serving
Soham Sharma21cabbe -
Serve Space frontend from root path
Soham Sharmacfacc82 -
Enforce exclusive reward schema bounds
1d6e12b -
Clamp reward metrics to open interval
34d60ee
-
Enforce open-interval task scores
6278f4a -
fix(inference): add structured output blocks for validator
7ef7162
-
Update data files: fto_profiles, incident_reports, resource_scenarios
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>2f65d5a
-
Pass multi-mode deployment validation: fix script key, health status, add missing endpoints
- pyproject.toml: rename script key to 'server' (checker requires exact name) - app.py: /health returns 'healthy' instead of 'ok' - app.py: add /metadata, /schema, /mcp endpoints required by runtime validator Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>4ea8201 -
Fix scripts entry point: add serve alias, include server package in setuptools discovery
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c83b427 -
Remove test and unused training files
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f537c14 -
Fix server entry point: add main() function and __main__ guard to server/app.py
- server/app.py: add callable main() that runs uvicorn - server/app.py: add if __name__ == '__main__' guard - pyproject.toml: point scripts entry to main, not app object Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c6af75e -
Fix provider-aware model pool: stop rotating to OpenRouter names on OpenAI/HF endpoints
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>7cedc7e -
Fix multi-mode deployment: uv.lock, server/app.py, scripts entry point, openenv-core dep
- Add uv.lock (118 packages resolved) - Add server/app.py entry point required by multi-mode deployment checker - Add [project.scripts] avigilance-server entry in pyproject.toml - Add openenv-core>=0.2.0 dependency - Bump openai to >=2.7.2 (required by openenv-core) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>3f99e49 -
Add pyproject.toml for multi-mode deployment readiness
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a68c050 -
Noisy imbalanced data: real-world messy conditions for harder agent evaluation
FTO data (1000 profiles): - DGCA-realistic distribution: 65% C, 22% B, 10% A, 3% A+ - Conflicting signals: good pass_rate/compliance but dangerous safety record - Near-boundary scores hovering at grade thresholds (±3 pts) - Ghost FTOs: 0 students/aircraft, 0 incidents (no flights = no safety record) Incident data (8000 reports): - Imbalanced severity: 40% low, 35% medium, 18% high, 7% critical (rare) - Extreme recurrence: 25% zero, 5% chronic (13-25 occurrences) - Deceptive resolved flags: critical incidents marked resolved (79 cases) - Noisy descriptions that contradict incident type - Stale inspections: 15% have >500 days since last check Resource scenarios (150): - Always underfunded: 1-4 inspectors, budget covers only 30-60% of items - Forced prioritization — agent cannot just assign everything Baseline scores updated (gpt-4o-mini, seed=42): task1=0.80, task2=0.71, task3=0.85, mean=0.79 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>fb646c3 -
Fix task2: varied incident severities + exact grader formula in prompt
- generate_data.py: severity now varied (low/medium/high/critical) instead of hardcoded 'high', which caused all priority scores to tie at 1.0 making the correct ranking unknowable by the LLM (0/3 top-3 overlap) - inference.py task2 prompt: compute priority scores using exact grader formula and pass pre-ranked list; escalate threshold now matches grader (score >= 0.85) - Result: task2 0.40 → 0.90, mean score 0.69 → 0.87 - Updated baseline scores in openenv.yaml and README Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b56b7c9 -
Switch to OpenAI API; fix key routing to use HF_TOKEN as universal credential
Hackathon evaluators set HF_TOKEN as the API key regardless of provider. URL-based routing now resolves: OPENAI_API_KEY (openai.com) → HF_TOKEN → OPEN_ROUTER_API. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>0044b6d
-
Fix generate_data.py PermissionError: chown /app to user before USER switch
data/ was created as root; non-root user couldn't write JSON files to it. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>606b357 -
Fix Space pip install: upgrade pip, --prefer-binary, cap major versions
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>86f7ec5 -
Drop uvicorn[standard] extras — uvloop/watchfiles need Rust compiler not available on HF build image
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>2477a22 -
Loosen Space dep pins — use >= to get pre-built wheels on HF build image
Strict pins like numpy==1.26.4 may trigger source compilation if no matching wheel exists. Loose bounds let pip choose the latest wheel. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b3b291c -
Remove scipy from Space deps — no Fortran compiler on HF build image
scipy==1.13.0 requires gfortran to build from source and is not used anywhere in app.py, inference.py, or the environment server code. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f786caf -
Dockerfile: remove --only-binary flag, keep python:3.11-slim
3.11 has pre-built wheels for all deps so compilation is not needed. --only-binary caused pip to fail on packages without a matching wheel name. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>2778998 -
Fix OOM: use python:3.11-slim and prefer binary wheels in Dockerfile
python:3.14-slim has no pre-built scipy/numpy wheels, forcing compilation which OOMKills the HF free-tier build. 3.11 has full wheel coverage. --only-binary=:all: avoids source compilation; fallback to source if needed. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>9e773be -
Fix OOMKilled build: use requirements-space.txt (no torch) in Dockerfile
torch is ~2GB and kills the HF free-tier build. The Space only runs FastAPI + LLM inference — no torch needed. Local RL training still uses requirements.txt which keeps torch. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>13c87f1 -
Merge branch 'main' of https://huggingface.co/spaces/soham27/AvigilanceEnv
a55fc40 -
initial commit
Verified signature857671f -
Initial submission: Avigilance 2.0 — India Aviation Safety Monitoring OpenEnv
Three-task OpenEnv for DGCA aviation safety oversight: - Task 1: FTO Quality Scorer (mean=0.8190, 100 episodes) - Task 2: Incident Prioritiser (mean=0.5358, 100 episodes) - Task 3: Resource Allocator (mean=0.8380, 10 episodes) LLM agent with rolling memory and multi-model free fallback pool. Grounded in MGURM framework with P1/P2/P3 safety principles. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>fe073e2