Spaces:
Paused
Paused
Commit History
-
Simplify blog to focus on core safety problem.
Made-with: Cursor33a237a -
added blogs
c847bf3
-
Add live weight learning for scripted target simulator
- TargetModelSimulator: apply_update with class-wide learned biases, defense products wired into respond (input/output + meta scale). - Battle sample includes blue_input_defense for input-layer updates. - Control center: friendlier weight labels and updated empty-state copy. Made-with: Cursora1e3ee9 -
Fix advanced suite UX, subject updates, and Pyright for apply_update
- Run advanced scenarios as sequential start/run/state with long fetch timeouts and polling; avoid single hung run-advanced request. - Apply subject learning and record_episode for each advanced task; share _apply_subject_update_after_episode with run_battle. - Control center: telemetry episode reset, weight column/chart fallbacks, stopPolling(silent) between suite scenarios, getattr for type safety. Made-with: Cursor0ac3cde -
chore: ignore local .cursor and HF upload cache under evidence
Made-with: Cursor6e8c72a -
Upload folder using huggingface_hub
Verified signature9c39df0 -
docs: add cloud inference and linked project guides
Add the blog post, cloud inference walkthrough, and cross-links across README and system docs while excluding binary screenshots from Git push so Space deployment succeeds. Made-with: Cursorf5ef44a -
Update active defense policy snapshot and persist latest battle logs.
Capture current checkpoint delta and newly generated public battle traces for recent runtime sessions. Made-with: Cursora353b4b -
Fix real-time chart rendering during live battle runs.
Lazy-init Chart.js instances when graphs tab first opens (avoids 0x0 canvas from hidden container), preserve array references with splice/ length=0 so chart datasets stay linked, and resize charts on tab switch. Made-with: Cursor745bf08 -
Fix live battle telemetry streaming and grader metric visibility.
Run episodes without blocking API polling, append chart points for every unseen turn, and surface grader reasoning with numeric scores in real time. Made-with: Cursor8f04120 -
Expand battle telemetry plumbing and notebook run-mode support.
Align backend state/update reporting with richer logs while updating training notebooks and smoke flows for advanced/autonomous execution paths. Made-with: Cursor45601e3 -
Add interactive Logs Explorer tab for viewing raw files
Introduces a new "Logs Explorer" tab in the UI that dynamically fetches a directory tree of all training artifacts, application logs, baseline scores, and battle sessions from the backend (including persistent /data storage on HF Spaces). The user can click any node in the tree to view its raw contents directly within the dashboard or download it. Made-with: Cursor8e45c5a -
Add Episode Core Metrics graph and dynamic hover tooltips
Update the Metrics & Graphs tab to include a new unified "Episode Core Metrics" chart that plots violation, novelty, hallucination, red reward, and blue reward all on one curve. Also overhauled the Chart.js tooltip logic globally so hovering over any graph node displays the exact Task ID and Episode ID (session) running at that exact turn instead of just the turn number. Made-with: Cursoreb4c582 -
Redesign public UI with modern aesthetic and live Chart.js graphs
Replace the old primitive dashboard with a highly styled modern glassmorphism/neon interface. Add a secondary "Graphs" tab that renders real-time, animated metrics (Weight Delta, Loss Profile, Grader Scores, Holdout KPIs) using Chart.js. Expand the HTML5 canvas battle scene with richer animations, particles, skybox, and shadow depth. Made-with: Cursore193d46 -
Fix Dynamic situations
f405c53 -
Add full transparency auditing, hallucinations, and citation checks
Expand the episode trace models to track all latent variables, reasoning blocks, and generated outputs. Update the UI to render citations, mock hallucinations, and explicit judge reasoning in real-time alongside the existing visualizer and scorecards. Made-with: Cursoreac3abc -
Improve battle reliability and upgrade realtime battle dashboard.
Harden non-scripted battle execution with safer LLM-subject fallback handling, and redesign the public UI to show richer auto-populated telemetry during battle runs. Made-with: Cursor2d543ed -
Fix Dynamic situations
38994b7 -
Fix Docker install permissions by installing deps as root
f7ddcfc -
Serve public backend UI, remove frontend build from Docker, expose public logs
34f8e25 -
Fix HF Docker build order for editable install
43d1976 -
Fix frontend lockfile for HF Docker build
32c5865 -
Prune local .cursor scaffolding and update README submission checklist.
This records the current workspace state by removing the previously committed .cursor toolkit files and keeping the README checklist aligned with current submission artifacts. Made-with: Cursorc2cde81 -
Stabilize runtime integrations and expand judge-focused project specification.
This unifies OpenEnv typing/runtime compatibility, fixes backend routing and registry endpoints, hardens training rollback/loading behavior, improves TUI event handling, and rewrites the README into a complete hackathon-facing project spec with judging alignment, training/evidence paths, and submission checklist. Made-with: Cursord4b35cb -
Implement layered multi-agent backend execution and reward/training integration.
This updates red/blue orchestration, battle execution, simulator and reward logic, and trainer/hybrid-learning plumbing so backend episodes and telemetry align with the new two-layer guardrail flow. Made-with: Cursor58d49be -
Add OpenEnv HTTP adapter and HF-ready notebook/deploy flow.
This makes the environment runnable on Hugging Face Spaces with compliant /reset and /step behavior, adds a Space deploy script, and updates Colab notebooks to run with scripted fallbacks when Gemini keys are missing. Made-with: Cursor5850cd5 -
Add system overview documentation and project pyright config.
Document the end-to-end architecture, training impact, and public commitments in a dedicated markdown guide, and add pyright configuration so imports resolve consistently against the project virtual environment. Made-with: Cursor508cfbf -
Add generated runtime logs and training artifacts snapshot.
Capture the latest end-to-end run outputs, including component-level logs and per-attack training artifacts, so the frontend/training demo state is reproducible from this revision. Made-with: Cursora9f2fbe -
Implement dynamic per-attack training with full observability.
Switch Red/Blue/Judge to provider-driven runtime decisions, add structured per-component run logs, and upgrade all Colab notebooks to execute end-to-end data generation, training, evaluation, and optional Hub publishing for the OpenEnv workflow. Made-with: Cursorc8cd24f -
Add hybrid learning lifecycle scaffolding for battle sessions.
Record per-episode runtime metrics, expose hybrid learning state in /api/metrics, and add rollback support so serving model versioning can be gated and observable. Made-with: Cursor66acf0c -
Expand GOAT-Evolution into a full end-to-end OpenEnv build.
Implement the multi-phase architecture with deterministic core modules, provider-aware agents, orchestration/session flows, OpenEnv-compatible server composition, cinematic React surfaces, TUI/training scaffolds, and local-first packaging so the project is runnable, testable, and demo-ready. Made-with: Cursor48f3216 -
Initialize GOAT-Evolution baseline with deterministic backend, battle APIs, and live frontend arena.
This establishes a runnable local foundation with tests, streaming battle events, and packaging/dev scaffolding to start executing the full build plan phases. Made-with: Cursor8da0e27