Skip to content

Releases

Mancelle paper evaluation results

Choose a tag to compare

@d0rbu d0rbu released this 21 Nov 04:27

Results and figure inputs for the revised Mancelle-3B checkpoint (step 1120, 20% drills retained, 50% FP self-conditioning dropout).

Download and extract mancelle-evaluations.tar.gz. The archive checksum is available alongside it.

  • full-suite/README.md explains how to reproduce the figures for ChatCORE, clean steering, calibrated counter-steering, GEPA, soft prompts, RL red-teaming, and negated-positive steering.
  • audit/FINAL_AUDIT.json records evaluation coverage and ChatCORE task counts.
  • calibrated-counter/calibrated-counter-instruction.pdf includes recovered instruction ratings for the magnitude-zero controls.
  • negative-steering/ contains the updated appendix figure, source tables, and plotting script for dedicated suppression and the revised Pythia baseline.

Mancelle-3B SFT dataset

Choose a tag to compare

@d0rbu d0rbu released this 21 Nov 04:48

The supervised fine-tuning mixture for the public Mancelle-3B step-1120 checkpoint. The training mixture retains 20% of the MC-binding drills (36,000 examples) alongside the instruction and synthetic steering examples.

  • train.jsonl.zst contains the conversations in Zstandard-compressed JSON Lines format. Extract with zstd -d train.jsonl.zst.
  • train.manifest.json records the mixture, derivation and checksums. Per-source counts refer to the original training mixture.
  • SHA256.json gives the checksums of the distributed files.

The release contains 1,246,590 rows. Three rows containing private-key headers were omitted and seven credential-shaped strings were redacted across three other rows. The order of the remaining rows is preserved.