Spaces:
Sleeping
Sleeping
Commit History
-
Add 02 / WHAT THE AGENT REVIEWS — three real env scenes
Pulls actual content from dep_010 (LLM-hallucinated packages), iac_015 (Terraform RDS misconfigs), migration_007 (CLUSTER on 4B-row table during zero-downtime deploy). Three side-by-side cards on desktop, stacked on mobile. Sells the env's innovation by showing real code + real findings rather than abstract architecture. Section indexes renumbered: 02 BENCHMARK→03, 03 RESOURCES→04, 04 INTERFACE→05. Nav reordered: Results · Scenes · Tasks · Resources. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>4233bf4 -
Expand BLOG.md to full end-to-end submission writeup
Comprehensive coverage assuming judges read only the blog: problem framing, env architecture, action space, scenario anatomy, semantic grader internals, SFT→GRPO hybrid pipeline, full hyperparams, per-task results with embedded plots, top-5 wins per task, training loss curves, what broke and what we learned, "what we shipped beyond v1" items as completed, full resource index, 60-second curl quickstart. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>5b5a584 -
Repoint blog links to BLOG.md file (per hackathon guidance)
Organizer clarified: prefer BLOG.md as a separate MD file at the HF Space root, not a Discussion or HF Pro blog. README and landing page now link to https://huggingface.co/spaces/sam25kat/securereview/blob/main/BLOG.md as the canonical writeup; the Community discussion remains as a mirror. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>c0449da -
Restructure landing for storytelling — proof-first
- Reorder: Hero → Stats → 01 RESULTS → 02 BENCHMARK → Thesis → 03 RESOURCES → 04 INTERFACE - Tighten hero subtitle to a one-line problem-frame ("AI writes the code. Who reviews it?") - Replace 0.98 oracle stat with +0.24 mean lift (story over compliance) - Stats card spacing: 160px → 80px - Section spacing: 140px → 80px - Hero padding: 140px → 80px - Nav padding: 120px → 80px - Add 6-line curl quickstart inline under interface section - Headline result cards now show baseline → trained inline Result: judges land, see the win in <2 scrolls. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>8602b50 -
Add Results + Resources sections to landing page
- Top nav now links to Tasks · Results · Resources · API · Blog · GitHub - Hero CTA primary button is now "View Training Results" - 04 / RESULTS section: headline cards (+0.302 / +0.295 / +0.126) + 3 inline plots - 05 / RESOURCES section: blog · RESULTS.md · SCENARIOS.md · plots dir · 3 trainers · GitHub - Footer expanded with Results + Blog links - All links open the public HF / GitHub URLs Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>e47ffb7 -
Add hackathon + team credit (Meta × Hugging Face OpenEnv · ~The Cook House)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>1ccf723 -
Update stats: 16 → 76 scenarios, 72 → 430 findings
Reflects the expanded scenario library across all three domains: dependency 6→24, iac 6→24, migration 4→28. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>76295c0 -
Link published HF Community blog from README
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>40d1711 -
Add SFT→GRPO hybrid pipeline, 60+ scenarios, semantic graders, full results
- Three-domain training results (dep +0.302, migration +0.295, iac +0.126) - Expanded scenario library: 24 dep, 24 iac, 28 migration = 76 hand-curated scenarios - Semantic-similarity graders across all three domains (alias dictionaries) - Hybrid SFT-warmup → GRPO-refinement training pipeline wired to live env - Multi-scale model study: Qwen 1.5B / 7B / 14B - Plots committed at training_results/plots/{dep,migration,iac}/ - Complete scenario index at training_results/SCENARIOS.md - Mini-blog draft at BLOG.md for HF community post Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>d2a68fa
-
Add HF training Space: Gradio UI + GRPO train script
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>443f900 -
Fix GRPOConfig: max_new_tokens → max_completion_length, remove temperature
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a28dc6a -
Add adaptive curriculum, GRPO training notebook, and beginner guide
- Adaptive difficulty: environment now tracks score history and auto-selects task/scenario tier based on rolling average (RLVE concept) - New GET /curriculum endpoint returns skill level, recommended task, progress - ResetRequest gains optional adaptive=true field - train_grpo.ipynb: 14-cell Colab notebook (Unsloth + TRL GRPO) training Qwen2.5-1.5B-Instruct against live SecureReview environment - GUIDE.md: comprehensive beginner guide for pitching and mentoring prep - plots/: directory for training artifacts (reward_curve.png, before_after.png) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>4c557cd
-
Refine product positioning: premium landing page + README
sameerkatte +1: Claude Opus 4.6 (1M context)Elevate SecureReview from 'a hackathon submission' to 'the agent review benchmark' with tighter copy, confident product language, and a minimalistic aesthetic inspired by Linear / Vercel / Raycast. Landing page (app/landing.py): - Extract inline HTML from main.py into dedicated landing module - Monochrome palette with single warm accent (orange) - Typography: Inter 200-weight hero + JetBrains Mono technical elements - Ambient radial glow, subtle grain overlay - New Thesis section framing the core argument - Four-stat quantitative bar (3 domains, 16 scenarios, 72 vulns, 0.98 oracle) - Formal domain titles (Dependency & Supply Chain Security, etc.) - Refined section headers (01/BENCHMARK, 02/INTERFACE, 03/QUICKSTART) - Responsive layout with thesis, task, and endpoint breakpoints README.md: - Hero with italic tagline and 'Security review, for the age of AI.' - New Thesis blockquote making the core claim concrete - Three-column benchmark table with formal domain names - Comparison table…6c30cc3 -
Add HTML landing page at GET / for HF Space preview
sameerkatte +1: Claude Opus 4.6 (1M context)The Hugging Face Space preview was showing a raw {"detail":"Not Found"} because no root route was defined. This adds a styled HTML landing page describing the environment, tasks, endpoints, and quick-start curl examples — much friendlier for anyone landing on the Space URL. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>ab388ea -
Clamp grader scores strictly within (0, 1) for validator compliance
sameerkatte +1: Claude Opus 4.6 (1M context)The OpenEnv Phase 2 Task Validation requires each task's score to be strictly between 0 and 1 — not exactly 0.0 and not exactly 1.0. Previously the grader could emit 0.0 (empty episode, max steps used) or 1.0 (perfect findings with maximum bonuses). Changes: - Reduce F1 coefficient from 0.85 to 0.83 so the maximum possible score is 0.99 instead of 1.00 (0.83 + 0.10 severity + 0.05 efficiency + 0.01 participation = 0.99) - Add a 0.01 participation bonus so empty episodes never score exactly 0.0 - Clamp final score to [0.01, 0.99] after rounding - Verified across all 16 scenarios in 4 edge cases (empty, perfect, immediate complete, all false positives) — every score falls strictly in (0, 1). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>382a35d -
Add [START]/[STEP]/[END] structured output markers to inference.py
sameerkatte +1: Claude Opus 4.6 (1M context)The validator parses stdout for these markers to extract per-task scores. Previously inference.py only printed freeform text, so the Output Parsing check failed with "No [START]/[STEP]/[END] in stdout". Changes: - Import sys/functools and wrap print() with flush=True so all output is flushed immediately (no buffering inside the validator harness) - run_episode() emits [START] task=NAME on entry - Each step emits [STEP] step=N reward=VALUE - Episode end emits [END] task=NAME score=VALUE steps=N Verified locally against the live HF Space: all 3 tasks emit the full marker set even when the LLM call fails (fallback to mark_complete). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>6dbb8cf
-
Pass openenv validate: add multi-mode deployment + runtime endpoints
sameerkatte +1: Claude Opus 4.6 (1M context)Local validator (openenv validate .): - Add pyproject.toml with openenv-core>=0.2.0 dependency - Add server/app.py with main() entry point - Add uv.lock for reproducible builds - Update openenv.yaml to spec_version: 1 format - Result: [OK] Ready for multi-mode deployment (docker, openenv_serve, uv_run, python_module) Runtime validator (openenv validate --url): - Fix /health to return {"status": "healthy"} (was "ok") - Add /metadata endpoint (name, description, version, tasks) - Add /schema endpoint (action, observation, state JSON schemas) - Add /mcp endpoint (minimal JSON-RPC 2.0 shim with tools/list) - Result: 6/6 criteria pass Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>8d618ab -
Fix /reset to accept empty/missing body for validator compatibility
sameerkatte +1: Claude Opus 4.6 (1M context)The OpenEnv validator calls POST /reset with no body, which was failing with 422 (missing body field). This change makes the ResetRequest body optional and defaults task_id to 'dependency_review' when omitted. - ResetRequest.task_id is now Optional[str] - /reset endpoint accepts Optional[ResetRequest] (None when no body sent) - Falls back to DEFAULT_TASK_ID when task_id is missing Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>ee2d45e -
Deploy to HF Spaces: add baseline scores, fix deps and validation
sameerkatte +1: Claude Opus 4.6 (1M context)- Deployed live on https://huggingface.co/spaces/sam25kat/securereview - Added README frontmatter for HF Spaces Docker SDK - Real baseline scores with DeepSeek-V3: dep=0.45, iac=0.52, mig=0.05 - Added openai/requests to requirements.txt (for inference.py) - Fixed HF_TOKEN validation warning in inference.py - Added operation type validation in migration grader - Dockerfile pip upgrade Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>2ee6649
-
Initial commit: SecureReview OpenEnv environment
sameerkatte +1: Claude Opus 4.6 (1M context)Complete implementation of AI security code review environment with: - 3 tasks: dependency review (easy), IaC review (medium), migration review (hard) - 16 hand-crafted scenarios with 72 total planted security findings - FastAPI server with /reset, /step, /state, /tasks, /health endpoints - F1-based grading with severity bonus, efficiency bonus, FP penalty - Baseline inference agent using OpenAI client - Docker support and openenv.yaml spec compliance Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>8b4c1a6