LASR-Callum/qwen3.6-27b-lora-500k-da20-t1t3

Qwen3.6-27B — 500k, 20% difficult-advice from traits 1-3 only

LoRA adapter trained on 501,212 tokens with loss on assistant tokens only, empty-think markers excluded from the loss, for 1 epoch at lr 4e-5.

Training data: qwen3.6-27b-mixture-500k-da20-t1t3.

Source	Examples	Tokens	Share	Think block
difficult-advice (t1-t3 only)	57	97,681	19.49%	real reasoning trace
NuminaMath-CoT	492	269,451	53.76%	none
No Robots	215	67,345	13.44%	empty marker
TULU3	94	66,735	13.31%	empty marker
Total	858	501,212

Within the non-difficult-advice 80.5%: NuminaMath 66.8%, TULU3 + No Robots 33.2%.

Trait coverage: the first three principles only

The difficult-advice half draws 19 examples each from t1, t2 and t3, and nothing from t4-t8:

	Principle	Included
t1	Honesty and non-deception	yes
t2	Respect legitimate oversight and norms	yes
t3	Avoid facilitating harm or illegality	yes
t4	Respect human autonomy	no
t5	Proportionate, non-preachy tone	no
t6	Genuine helpfulness within ethical bounds	no
t7	Nuance over rule-following	no
t8	Prioritize the long-term good	no

The three included principles are the prohibition-shaped ones; the five excluded lean toward tone and judgement.

The paired comparison

The non-difficult-advice half is byte-identical to …-mixture-500k-da20-numina -- the same 492 NuminaMath, 215 No Robots and 94 TULU3 rows, with the same markers. The two datasets differ in exactly one respect: whether the difficult-advice covers all 8 principles (7 each) or only the first three (19 each). That isolates trait breadth from every other variable.

Think-block convention

Data	Renders as	In the loss?
difficult-advice	<think>real reasoning</think>	yes -- this is the signal
TULU3, No Robots	<think>\n\n</think>	no -- context only
NuminaMath-CoT	no block; its CoT is in the response text	n/a

The empty marker is masked from the loss: the model is conditioned on it but never trained to emit one, since learning to emit an empty think block is the documented reasoning-collapse pattern.

Training

Supervised	395,352 / 501,212 = 78.9%
Epochs / steps	1 / 54
lr / schedule	4e-5, cosine, 3% warmup
Runtime	40 min, 1x H100 80GB
r / alpha / dropout	32 / 64 / 0.05
batch x grad-accum	1 x 16
max seq len / packing	3072 / off
Final loss	0.952
Token accuracy	0.775

Verified before training, on the box: zero empty-think markers inside any supervised span, zero user or system tokens in the loss, and all 57 difficult-advice rows retaining their real reasoning traces.

The paired run

Run	Difficult-advice traits	Loss	Token acc
all 8 traits	t1-t8, 7 each	0.946	0.776
this	t1-t3 only, 19 each	0.952	0.775

Same seed, hyperparameters and non-difficult-advice half; only trait breadth differs. Training loss is nearly identical, which is expected -- both see a comparable volume of difficult-advice tokens. Whether narrow coverage generalises to the five unseen principles is a question for evaluation, not for training loss.

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

Usage

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/qwen3.6-27b-lora-500k-da20-t1t3")
model = model.merge_and_unload()

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