Downloads · 30 days
0
dougalldeepmind/2026-08-25-qwen36-lora-table2-9284-gpt-responder-685-paired-rank-64
2026-08-25-qwen36-lora-table2-9284-gpt-responder-685-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. The card lists the license as apache-2.0.
Third arm of the generator ablation. Trained on 685 difficult-advice rows whose assistant turns were drafted by openai/gpt-5.6-luna and revised by openai/gpt-5.6-terra, plus the shared 9,284-row Table2 half.
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
Third arm of the generator ablation. Trained on 685 difficult-advice rows whose assistant
turns were drafted by openai/gpt-5.6-luna and revised by openai/gpt-5.6-terra,
plus the shared 9,284-row Table2 half.
Its siblings answer the same 716 questions — same situations, user turns and system prompts, authored by Haiku/Sonnet and reused verbatim, selected by scenario id rather than sampled:
| arm | answers written by | adapter |
|---|---|---|
| A (baseline) | Haiku 4.5 → Sonnet 5 | LASR-Callum/2026-08-14-qwen36-lora-table2-9284-difficult-advice-716-rank-64-dynbatch |
| B | grok-4.6 (both stages) | LASR-Callum/2026-08-24-qwen36-lora-table2-9284-grok-responder-703-paired-rank-64 |
| C (this) | gpt-5.6-luna → gpt-5.6-terra | this repo |
checkpoint-600 of 624 steps, not a final adapterThe run crashed on its last step. The mixture is 9,969 rows; at global batch 16 that is
623 full steps plus a final batch of exactly one example, and route_step cannot split
one example across two DDP ranks:
ValueError: route_step cannot give all 2 ranks work from 1 examples in 1 passes
The trainer therefore never wrote its own adapter/ directory, and training_meta.json
here was backfilled from the train config rather than emitted by the trainer.
Why this is judged immaterial, so you can disagree with the judgement rather than discover
it: the cosine schedule had already collapsed. LR at step 600 was 4.4e-07, and the 24
missing steps ran from there down to 1.8e-08 — 200x to 5,000x below the 1e-4 peak. The
parameter movement in that window is a rounding error next to the other differences between
these arms (response length varies 1.57x / 0.59x across them). checkpoint-600 is 96.2% of
one epoch.
If you want the clean number, checkpoint-600 in the training output carried optimizer.pt
and scheduler.pt, so a single-GPU resume of the last 24 steps would finish it — on one
rank the 1-example batch is fine, because route_step only fails when it must split.
Sibling arms B and A completed their epochs. Their mixtures happened to have remainders that split cleanly (arm B: 9,987 rows, remainder 3). This is a row-count edge case, not a property of the data.
| reasoning median | response median | vs baseline | |
|---|---|---|---|
| baseline (Haiku→Sonnet) | 2,876 | 2,670 | — |
| arm B (grok) | 1,277 | 1,568 | 0.59x shorter |
| arm C (this, GPT) | 2,114 | 4,195 | 1.57x LONGER |
So the sweep brackets the baseline rather than only undershooting it. Any behavioural difference between these arms is confounded with response length — report length alongside every downstream metric.
Two deviations arm B could not avoid are avoided here:
reasoning: {enabled: false}; grok-4.6 and Gemini 3.x return 400 on it and had to run at
effort: low, meaning their corpora were produced by a different computation than the
baseline's non-thinking Haiku/Sonnet. This arm matches the baseline regime exactly.In draft_responses only: retries 5 (baseline 2) and max_fail_pct 10.0 (baseline 2.0),
because ~20% of luna's single-shot drafts land under the stage's 700-char floor. min_chars,
all 13 ban_patterns and every prompt are unchanged. 31 of 716 rows were lost, all at the
draft stage; terra lost zero.
LASR-Callum/2026-08-25-difficult-advice-gpt-responder-716LASR-Callum/2026-08-25-table2-9284-gpt-responder-685-paired-train-mixtureLASR-Callum/2026-08-25-gpt-responder-685-paired-bundleconfigs/train/2026-08-24_lora_qwen36_table2_9284_gpt_responder_685_paired.yamlthinking: true