matboz's picture
code bundle @ e3267a68
7e24f6c verified
|
Raw
History Blame Contribute Delete
1.16 kB
metadata
license: apache-2.0
tags:
  - sft
  - lora
  - qwen3.6
  - training-bundle

Training bundle — 4xH200 LoRA run

code.tar.gz only: the trainer, src/ and configs/. The mixture is pulled separately from LASR-Callum/2026-08-08-table2-9000-synthdoc-1000-traitbalanced-len8000-train.

field value
experiment Code bundle for a 1-epoch LoRA SFT of Qwen/Qwen3.6-27B; see the mixture repo for composition
date_generated 2026-08-07
constitution claude_distilled_09_principles_mid_20260804 — via the mixture's 20% side; see the mixture card
source_repo teaching_claude_why_replication @ e3267a68053048dfc188e76f37958b90d0de92a5
models trains Qwen/Qwen3.6-27B
generation_config n/a — this repo carries code, not generated data
schema code.tar.gz: scripts/train/train_lora.py, src/, configs/
provenance uv run python scratch/launch_memself_h200x4.py bundle

Training config inside the tarball: configs/train/lora_qwen36_t2_9000_synthdoc_1000.yaml.