Assistant Axis LoRA weight projection experiment
Research outputs, ordinary PEFT adapters, judged responses, and timing logs. Source and target base weights are not included.
Source: mistralai/Mistral-7B-Instruct-v0.3. Target: meta-llama/Llama-3.1-8B-Instruct. Teacher and judge: Qwen/Qwen3-14B / Qwen/Qwen3-14B.
Assistant response preferences only; no biographies or personal factual memories. The benchmark and weight projection are experimental, not published benchmark reproductions.
Dolly questions retain CC-BY-SA-3.0 attribution to Databricks; PersonaAtlas profile fields come from HowieHwong/PPOpt-data under Apache-2.0. Adapters require their respective base model and are subject to the terms of the named official base-model repositories.
See provenance.json for pinned revisions, data selection, and exact runtime versions. Results are in metrics/summary.json and metrics/report.md after completion.
- Downloads last month
- -