Upload README.md with huggingface_hub

README.md

@@ -32,10 +32,11 @@ This provides a natural signal for training models to:
 
 ### Key Features
 
-- **128K+ positive triplets** from 617 top GitHub repositories
-- **40K+ negative examples** (~24% of dataset) of clean code labeled "No issues found."
-- **36 programming languages** (Python, TypeScript, Go, Rust, C++, JavaScript, C#, Java, Kotlin, Swift, and more)
-- **Quality-filtered**: bot comments, noise ("LGTM", "+1"), and auto-generated content removed
+- **167K+ positive triplets** from 725 top GitHub repositories
+- **51K+ negative examples** (~23% of dataset) of clean code labeled "No issues found."
+- **37 programming languages** (Python, TypeScript, Go, Rust, C++, JavaScript, C#, Java, Kotlin, Swift, and more)
+- **Human-only reviews**: AI/bot reviewers (Copilot, linter bots, etc.) are excluded
+- **Quality-filtered**: noise ("LGTM", "+1"), and auto-generated content removed
 - **Chunk-focused**: ~50 lines of context around the reviewed code, not entire files
 - **Permissive licenses only**: all source repos use MIT, Apache-2.0, BSD, or similar licenses
 - **Verified changes**: only includes triplets where the code chunk actually changed after the review
@@ -115,7 +116,7 @@ negatives = ds["train"].filter(lambda x: x["is_negative"])
 
 ### Quality Filters Applied
 
-- Bot authors excluded (dependabot, renovate, codecov, etc.)
+- Bot and AI reviewers excluded (dependabot, Copilot, golangcibot, houndci-bot, etc.)
 - Comments < 30 characters excluded
 - Noise patterns excluded (LGTM, +1, emoji-only, etc.)
 - Auto-generated comments excluded (coverage reports, CI output)