Views
No views yet
| field | value |
|---|---|
experiment | LoRA SFT adapter — lora_qwen36_less_top10_220_r64. THE LESS ARM: base Qwen3.6-27B trained on the 220 highest-influence rows (top 10%) of the difficult-advice pool by LESS score_max. The two arms differ only in which 220 rows they saw; every hyperparameter, the batching protocol and the loss path are identical, so a difference between them is attributable to the selection. Trained one arm per GPU on a single pod so both share one base-model download. |
date_generated | 2026-08-19 |
constitution | constitutions/claude_distilled_12_principles_mid/constitution.md — inherited from the training data (LASR-Callum/2026-08-19-less-top10-difficult-advice-220-train) |
source_repo | https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git @ 4078304ba62de85cca9e4c1d9347c4d88c95d52b |
models | base: Qwen/Qwen3.6-27B |
generation_config | {"seed": 0, "thinking": true, "epochs": 4, "lr": 0.0001, "batch_size": 1, "grad_accum": 16, "max_seq_len": 8192, "lora": {"r": 64, "alpha": 128, "dropout": 0.05}} |
schema | PEFT LoRA adapter (safetensors) + tokenizer + training_meta.json {thinking, train_config, base_model, dataset{repo,file,revision}, git_sha, timestamp} |
provenance | uv run train --config configs/train/2026-08-19_lora_qwen36_less_top_10_220_rank64.yaml push=false (on pod), then uv run python scratch/less/train_arms.py push |
dataset | hf.co/datasets/LASR-Callum/2026-08-19-less-top10-difficult-advice-220-train@9cb349966b2550aa20e58bf5f313a5160a31c1f2 (mixture.jsonl) |
paired_arm | LASR-Callum/2026-08-19-qwen36-lora-random-220-control-rank-64 |