Downloads · 30 days
0
dougalldeepmind/2026-08-26-qwen36-lora-table2-9284-sonnet-concise-703-paired-rank-64
2026-08-26-qwen36-lora-table2-9284-sonnet-concise-703-paired-rank-64 is a machine learning model from dougalldeepmind. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Downloads · 30 days
0
Access
Public
Updated Aug 31, 2026
Repo size
1.3 GB
Likes
0
Public
Click a slice to open those files.
.safetensors1.3 GB · 98%
From the Hugging Face model README
| field | value |
|---|---|
experiment | LoRA SFT adapter -- lora_qwen36_t2_9284_sonnetconcise703_paired_2xh200. ARM C, the LENGTH CONTROL of the generator ablation: Qwen3.6-27B on 9,284 Table-2 rows + 703 difficult-advice rows whose assistant turn is the baseline's own Haiku draft rewritten by Sonnet 5 under a one-sentence cap at grok's median lengths (reasoning ~220 words, reply ~270). Same 703 questions, same Table2 half, same seed/LoRA/schedule as arm B (LASR-Callum/2026-08-24-qwen36-lora-table2-9284-grok-responder-703-paired-rank-64); differs from the da716 baseline only in the rewrite's length. Read C's ODCV against A (16.3%) and B (7.8%): near B means length carried the drop, near A means the generator did. |
date_generated | 20260826_1 |
constitution | constitutions/claude_distilled_12_principles_mid/constitution.md (sha fe2ed96093d68a87...), inherited from LASR-Callum/2026-08-26-table2-9284-sonnet-concise-703-paired-train |
source_repo | https://github.com/Matthew-Bozoukov/Lessons_from_constituitional_AFT.git @ 8b9868cb3f1ea6ddb378cf659b65958f4cc6f702 |
models | base: Qwen/Qwen3.6-27B; synth half written by anthropic/claude-haiku-4.5 (draft) + anthropic/claude-sonnet-5 (length-capped rewrite) |
generation_config | {"seed": 0, "thinking": true, "epochs": 1, "lr": 0.0001, "batch_size": 1, "grad_accum": 16, "max_seq_len": 8192, "dynamic_batching": "token budget from ModelProfile.train_memory (H200 8000)", "ddp_ranks": 2, "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 | torchrun --nproc_per_node=2 scripts/train/train_lora.py --config configs/train/2026-08-24_lora_qwen36_table2_9284_sonnet_concise_703_paired.yaml push=false (on a RunPod 2xH200 pod via scripts/gpu/runpod_train.py), then uv run python scratch/sonnet_concise/adapter_pod.py push |
dataset | hf.co/datasets/LASR-Callum/2026-08-26-table2-9284-sonnet-concise-703-paired-train@9289a85eda51ea3341e2428e5b0d693947437cd9 (t2_9284_sonnetconcise703_10k.jsonl) |
comparison_arms | A: LASR-Callum/2026-08-14-qwen36-lora-table2-9284-difficult-advice-716-rank-64-dynbatch; B: LASR-Callum/2026-08-24-qwen36-lora-table2-9284-grok-responder-703-paired-rank-64; D: LASR-Callum/2026-08-25-qwen36-lora-table2-9284-gpt-responder-685-paired-rank-64 |