Spaces:
Sleeping
Sleeping
Commit History
-
Update BLOG.md
Verified signaturef56769c -
Update BLOG.md
Verified signaturea5ca786 -
Update README.md
Verified signatureaf4c96c -
Update README.md
Verified signature0ed8b20 -
Update README.md
Verified signature65e1955 -
Update README.md
Verified signaturedc614a9 -
Update README.md
Verified signaturee932bc3 -
Update BLOG.md
Verified signatureaaeff51 -
Update BLOG.md
Verified signaturebd38765 -
Update BLOG.md
Verified signature737ae15 -
Update BLOG.md
Verified signature3485a2d -
Update BLOG.md
Verified signaturee8d0a53 -
Update README.md
Verified signature49227b4 -
Update README.md
Verified signature7ec5eb4 -
Update BLOG.md
Verified signature3024a77 -
Update BLOG.md
Verified signaturebd79fcb -
Update BLOG.md
Verified signature3a7c922 -
Update BLOG.md
Verified signature72023f9 -
Update BLOG.md
Verified signature0f84b75 -
Update BLOG.md
Verified signature88df554 -
Update BLOG.md
Verified signature2e403bc -
Update BLOG.md
Verified signature76d9815 -
Update BLOG.md
Verified signaturea963cbd -
Update BLOG.md
Verified signature4861e01 -
Update BLOG.md
Verified signatureb780c1c -
Update BLOG.md
Verified signature87978ef -
Update BLOG.md
Verified signaturec78ca66 -
Update openenv.yaml
Verified signature54a7c04 -
Fix all submission gaps: openenv.yaml, README captions, baseline table, blog link
- Add long_horizon, personalized, curriculum tasks to openenv.yaml - Update openenv.yaml description and tags - Add before/after baseline comparison table to README - Add one-line captions under all 3 training curve images - Add BLOG.md link in All Links section - Fix Blog.md typo → BLOG.md in repo structure Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>fcd74c3 -
Remove frontmatter from BLOG.md
3d2e6c1 -
Update BLOG.md
233383e -
changes to blog
8d4f07a
-
pushed changes to readme
f11d97b -
update read
052c301 -
Update readme.md
c60cdcb -
Update README.md
d537044 -
Rewrite README: blog-style with animations, 10 tasks, new endpoints
3391ffe -
Update BLOG.md and README.md
bbe2575 -
Fix notebook widget state key placement (inside widget-state+json)
95d36d4 -
Remove BLOG_DRAFT.md and ROUND2_PROBLEM_STATEMENT.md
f115995 -
Maximize environment: curriculum task, metrics endpoint, 5 bug fixes, notebook fix
- Add /curriculum task: auto-progresses easy→medium→hard→expert based on score - Add /metrics endpoint: per-task episode counts, avg/best scores, regulator state - Fix /grader for long_horizon and personalized (was falling to expert grader) - Fix _PERF_HISTORY to track long_horizon, personalized, curriculum - Fix long_horizon/personalized/curriculum done condition (no early exit at 0.95) - Fix long_horizon phase 3: audits same invoices as phase 1 (not a separate batch) - Fix notebook widget metadata: add missing 'state' key to all 3 notebooks Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>4890422 -
Update all 3 GRPO training notebooks (tokens removed)
a3ecb7e -
Add long_horizon/personalized tasks + GitHub-hosted training curves
Training curves: - Drop matplotlib rendering; serve real Colab PNGs from GitHub raw URLs via gr.HTML (images hosted at github.com/ps2181/invoice-processing-pipeline) New tasks: - long_horizon: 20-step 4-phase episode (extract→reconcile→audit→forecast) sparse rewards, progressive reference data unlock per phase - personalized: 10-step adaptive task tracking per-field accuracy (vendor/date/math/completeness), generates next invoice targeting weakest field - New InvoiceObservation fields: phase, phase_context, agent_profile - New InvoiceState fields: phase, phase_scores, phase_context, agent_profile Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>473ab10 -
Implement all audit fixes: base class, curves, README, code quality
Core fixes: - environment.py: InvoiceEnvironment now inherits openenv_core.Environment with graceful fallback; SUPPORTS_CONCURRENT_SESSIONS=True; state property satisfies abstract method; super().__init__() called safely - web_ui.py: _make_training_curves() returns base64 HTML <img> with exact Colab log data (auditor peak 0.719 at step 10 with ±std band); gr.HTML instead of gr.Image/gr.Plot — renders reliably on HF Space Gradio 6 - client.py: episode_id param added to step() on both sync/async clients - app.py: /multi/approve now accepts JSON body (MultiApproveRequest) not query param; WebSocket reset wrapped in try/except; import random added - multi_agent_environment.py: fix _generate_expert_batch_biased return type hint (3→4 tuple) and docstring - pyproject.toml: remove duplicate python-dotenv dependency - __init__.py: relative import with fallback for package/script usage - BLOG renamed to BLOG.md README rewrite: - Mermaid 5-agent loop diagram - Before/after training r…b02956e -
Remove show_download_button — unsupported in this Gradio version
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>e422a3e -
Fix training curves: switch gr.Plot → gr.Image with PNG bytes
gr.Plot fails silently on HF Space with matplotlib figures. Save figure to PNG via BytesIO and render with gr.Image instead — PIL Image rendering is reliable regardless of Gradio version. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>7f1e860 -
Fix /web 404: guard matplotlib import and harden Gradio mount
- Move matplotlib/numpy imports inside _make_training_curves() so an import failure doesn't crash build_ui() at startup - Wrap _make_training_curves() call in try/except so chart failure falls back to gr.Plot(value=None) instead of killing the whole mount - Add traceback.print_exc() to Gradio mount except block so failures are visible in Space logs instead of silently swallowed - Add missing `import random` at top of app.py (generator/score endpoint called random.random() without importing it) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>707a8d9 -
Update training curves with real Colab data
Replaced all hardcoded/estimated values with actual training log numbers: - Extractor: 4-point curve, live grader peaked at 0.914 (step 15), crash visible - Auditor: run 1 flat (0.01 dead reward), run 2 real log values (0.28→0.52) - Generator: live evasion dead (0.01), fraud plausibility ~0.22 (format learned) - Summary panel: baseline vs best achieved per metric Tab description updated to honest table showing real outcomes per agent. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>5eea840 -
Add Training Results tab with GRPO reward curves for all 3 agents
- New tab "📈 Training Results" showing matplotlib plots (dark theme): - Extractor: run 1 crash at step 90 (MAX_SESSIONS bug), run 2 to 200 steps - Auditor: run 1 flat (episode_id bug), run 2 climbing to 0.35 - Generator: 0.10 → 0.77 adversarial reward over 50 steps - Summary bar chart: baseline vs final for all 3 agents - Plot rendered at startup via gr.Plot(value=_make_training_curves()) so it's visible immediately without any click Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>5a9c33c -
Overhaul UI + README for submission
web_ui.py: - Pipeline tab moved first — most impressive thing shown immediately - Full CSS overhaul: Inter font, dark monospace trace box, hero banner, agent card row, better button styling - Hero HTML + agent cards (icon, name, desc) shown on every tab - Pipeline trace: dark terminal style, ▓ bars, cleaner step headers, 🚨/✅ verdict icons, no "rule-based fallback" label - Regulator report: reformatted with consistent spacing and section headers - Tab names: "🔄 Live Pipeline", "📊 Regulator Dashboard", "🧪 Single Agent Tester" README.md: - Complete rewrite for HF Space landing page - Leads with the adversarial self-improvement story - 3 novel features table, all 3 trained LoRAs with HF Hub links - Full reward signal tables, API endpoint reference, fraud type table Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>ed15028