PasteWhat-Ranker-v1

Research is in progress. No trained, calibrated and accepted production model is published here yet.

The project distills decision labels into the original non-quantized Laya-multilingual encoder. At the user's request, future data generation and labeling use Pi's devin/swe-2; earlier valid kimi-for-coding data retain their original labels and provenance. New candidate and candidate-group-aware abstention heads score 1–20 existing clipboard entries. The model does not generate paste content or reasoning traces. Model inputs use application categories, not real app identities.

The teacher switch preserves 404 accepted episodes from the current production run (217 Train, 62 Dev, 58 Calibration, 67 Test). These are the counts at transition, not completed dataset quotas. Author, primary labeler and additional reviewer sources are recorded separately, including reused Kimi author drafts labeled later by SWE-2. The append-only transition policy binds the unchanged run plan and source/runtime hashes. One six-case Pi protocol check passed; it does not establish teacher or student benchmark accuracy.

The first local GPU engineering run is complete: 32 independently agent-reviewed Train examples were fitted in six epochs / 24 full-encoder updates, with all 32 training decisions correct. Conversion to MLX FP16 preserved those 32 decisions. This is training-set fit, not generalization accuracy, and that engineering checkpoint is not offered as a production model. Agent and teacher review is not human validation.

Remaining production work includes the 5k pilot, 20k full training with three seeds, new-pool hard-example training, independent final MLX verification, calibration and frozen paired Test. Planned counts and acceptance targets are not reported as completed experiments. Train/Dev production, training, and Calibration/Test evaluation are owned by different agents with conceptual-family separation.

The final bundle will include PyTorch reference weights, MLX FP16 deployment weights, tokenizer, preprocessing, calibration policy, training/data manifests and measured quality/performance reports. Any unmet target will be disclosed.

Source and ongoing execution records: GitHub. The AppKit application already supports Jev and has an adapter for the future calibrated local ranker. Synthetic-only metrics will not be claimed as real-user accuracy.

Apache-2.0; upstream source and conversion implementation attribution are documented in the source repository's LICENSE and NOTICE.