The zero-dose end of the synthdoc_v2 sweep. Same seed, hyperparameters, rendering and max_seq_len (2048) as the 10/90, 15/85 and 20/80 arms — only the difficult-advice share differs. It isolates what SFT on replay alone does, which a comparison against the base model cannot.

Training data: qwen3.6-27b-synthdocv2-mixture-0_100.


Tokens	995,877 (1,555 conversations)
Supervised	775,839 / 995,877 = 77.9%
Epochs / steps	1 / 98
Runtime	58 min, 1x H100 80GB
r / alpha / dropout	32 / 64 / 0.05
batch x grad-accum	1 x 16
lr / schedule	1e-4, cosine, 3% warmup
Final loss	0.876
Token accuracy	0.774

Loss runs slightly below the difficult-advice arms by construction: pure replay is a narrower target than a mixture carrying long reasoning traces. That is not a quality signal.

The sweep

Arm	Difficult-advice	Loss	Token acc
0/100 (this)	0%	0.876	0.774
10/90	10.0%	0.901	0.811
15/85	15.0%	0.924	0.768
20/80	20.1%	0.928	0.791

Not yet evaluated on ODCV-Bench or agentic-misalignment.

Usage

python

from peft import PeftModel
from transformers import AutoModelForImageTextToText

model = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.6-27B", dtype="bfloat16")
model = PeftModel.from_pretrained(model, "LASR-Callum/2026-08-02-qwen36-synthdoc-package-lora-0-100")
model = model.merge_and_unload()

Use AutoModelForImageTextToText, not AutoModelForCausalLM — this is a vision-language checkpoint.

LASR-Callum/qwen3.6-27b-synthdocv2-lora-0_100

Qwen3.6-27B — 0/100 zero-dose control (TULU3 only)

The sweep

Usage

LASR-Callum/
qwen3.6-27b-synthdocv2-lora-0_100