Conversation
|
Understand this PR’s impact Explore downstream dependencies and potential security impact with Blast Radius. No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Advanced Run ID: 📒 Files selected for processing (1)
Included review availability: Your plan provides up to 2 included reviews per hour; 1 remains after this review. 📝 WalkthroughWalkthroughChangesProject catalog
Priority: ⬇️ Low Estimated code review effort: 1 (Trivial) | ~3 minutes Change: Feature Suggested reviewers: 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
Landed on main as d155b64 with you as the commit author, now live on https://awesomejev.com. Closing rather than merging because every PR appends to the same data file and they conflict with each other; landing directly keeps your authorship. Thanks! |
Adds minojev to the
researchcategory (independent replica, one entry).I maintain the project. It is an independent Jev-style replica — no TypeSafe API calls — built around head training: the backbone stays frozen, candidate-path features are cached once, and only the small decision head is trained, followed by dev-fitted temperature calibration.
What it does: returns calibrated
Choice(2–255 candidates),Boolean(Noul-shaped), andScore(2–10 ordered levels) distributions in one forward pass withdecode_steps: 0. Code consumes the distribution and applies its own thresholds; the confidence explorer page demonstrates accept/escalate at a chosen confidence.Evidence committed in the repo:
results/compare-general.md— same-backbone head-to-head against token generation on a balanced 120-decision suite (95.8% vs 80.0% accuracy, 0 vs 3.48 output tokens/decision, ~5× better p95).results/general-metrics.json— calibrated 200-request evaluation: 97.5% accuracy, ECE 0.014, selective accuracy at ≥0.9 confidence: 89.5% coverage at 98.9% accuracy.Boundaries stated honestly: trained only on permissively licensed sources (MIT / Apache-2.0 / CC-BY); NC datasets are reserved for evaluation. On a deliberate out-of-distribution workload the head-trained model scored below a zero-shot generative baseline (31.7% vs 40.0%), which the repo reports rather than hides.
Links: repo · demos · benchmark page · models · data
This PR changes only
data/projects.json;npm run validatepasses (697 entries valid).Need help on this PR? Tag
@codesmith-botwith what you need. Autofix is disabled.Summary by CodeRabbit