Downloads · 30 days
0
dougalldeepmind/2026-08-28-gpt5-qwen36-lora-table2-9284-gpt-responder-685-paired-rank-64-seed42
2026-08-28-gpt5-qwen36-lora-table2-9284-gpt-responder-685-paired-rank-64-seed42 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_gptresp685_paired_s42_2xh200. Seed-42 REPLICATE of the GPT-responder paired arm (LASR-Callum/2026-08-25-qwen36-lora-table2-9284-gpt-responder-685-paired-rank-64 is seed 0): Qwen3.6-27B on 9,284 Table-2 rows + 685 difficult-advice rows whose assistant turn was drafted by openai/gpt-5.6-luna and revised by openai/gpt-5.6-terra. Identical data (byte-for-byte), code (seed 0's own bundle), LoRA shape, schedule and global batch; only the seed (LoRA init + shuffle order) differs. Purpose: the between-seed error on the arm's ODCV number. |
date_generated | 20260828_1 |
constitution | constitutions/claude_distilled_12_principles_mid/constitution.md, via the corpus LASR-Callum/2026-08-25-difficult-advice-gpt-responder-716 |
source_repo | https://github.com/Matthew-Bozoukov/Lessons_from_constituitional_AFT.git @ dd00d413b603a12eac63ab3ce1fc8c2caf58be58 (branch worktree-gpt-seeds) |
models | base: Qwen/Qwen3.6-27B; synth rows by openai/gpt-5.6-luna (draft) + openai/gpt-5.6-terra (revision); prompts by anthropic/claude-haiku-4.5 + anthropic/claude-sonnet-5 (the baseline's, reused verbatim) |
generation_config | {"seed": 42, "thinking": true, "epochs": 1, "lr": 0.0001, "lr_scheduler": "cosine", "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_args.bin + trainer_state.json (loss history) + training_meta.json {thinking, train_config, base_model, seed, dataset{repo,file,revision}, checkpoint{global_step,total_steps,epoch,lr_at_checkpoint,why_not_final}, git_sha, timestamp} |
provenance | uv run python scripts/gpu/runpod_train.py up --bundle LASR-Callum/2026-08-28-gpt-responder-685-seeds-bundle --train_config configs/train/lora_qwen36_t2_9284_gptresp685_paired_s42_2xh200.yaml --gpu 'NVIDIA H200' --gpu_count 2 --mixture t2_9284_gptresp685_10k.jsonl (torchrun --nproc_per_node=2 scripts/train/train_lora.py on the pod, push=false); scratch/low_stakes/train_watchdog.py pulled adapter.tar.gz; scratch/gpt_seeds/adapter_pod.py extract/push --seed 42 |
dataset | hf.co/datasets/LASR-Callum/2026-08-28-gpt-responder-685-seeds-bundle@9b1b30fdd2a621dfa24e21c90c0de4b5c4387dc6 (t2_9284_gptresp685_10k.jsonl; sha256-identical to LASR-Callum/2026-08-25-gpt-responder-685-paired-bundle@c7994ec2, seed 0's pin) |
checkpoint | checkpoint-600 of 624 steps (epoch 0.9615, lr 4.38e-07, last logged loss 0.883906078338623). run crashed at step 624/624: route_step cannot split a final 1-example batch across 2 DDP ranks (9,969 rows at global batch 16 leaves remainder 1). checkpoint-600 is 96.2% of the epoch -- the SAME end state as seed 0 (see its card), kept on purpose so the replicates share seed 0's protocol exactly. |
comparison_arms | seed 0: LASR-Callum/2026-08-25-qwen36-lora-table2-9284-gpt-responder-685-paired-rank-64; A (Sonnet): LASR-Callum/2026-08-14-qwen36-lora-table2-9284-difficult-advice-716-rank-64-dynbatch; B (grok): LASR-Callum/2026-08-24-qwen36-lora-table2-9284-grok-responder-703-paired-rank-64 |