Modelpublic

aamish-ahmad/behaviortune-v1-1-r1-adapter

sourceHugging Faceotherupdated 22d agoView on Hugging Face

0likes59downloads

Model Card

BehaviorTune QLoRA Adapter

QLoRA adapter for Qwen/Qwen3-4B-Instruct-2507, trained on the 240-row BehaviorTune training split and evaluated with matched BASE / SYSTEM / CONTEXT / QLoRA conditions.

What this demonstrates

—Open-model post-training with Transformers, TRL, PEFT, and QLoRA.
—Completion-only training on a controlled synthetic dataset.
—Deterministic matched evaluation against the untuned base model.
—Public adapter, dataset, results, and reproducible engineering tooling.

Measured result

On the frozen eval_core set, QLoRA increased format-valid activation from 0.65625 to 1.00000 — a +0.34375 (+34.375 percentage-point) matched shift, with a paired-bootstrap 95% CI of [0.1875, 0.5].

Training

Item	Value
Base model	`Qwen/Qwen3-4B-Instruct-2507`
Base revision	`cdbee75f17c01a7cc42f958dc650907174af0554`
Training rows	240
Epochs	3
Optimizer steps	90
Quantization	4-bit NF4, BF16 compute
LoRA	rank 32, alpha 64, dropout 0.05, all linear projections
Effective batch size	8
Learning rate	0.0002
Seed	147
Hardware	NVIDIA A100-SXM4-40GB

Project resources

Artifact identity

—Adapter SHA-256: 8d16ef2cb6ff7a982511fd58f21eff52538761f4d198b4cc5cbfd73ca7c9d4de
—Frozen run identifier: V1.1-R1
—Public release: BehaviorTune v1.0.0

Use and limitations

Load the pinned base revision and attach this PEFT adapter.

The measured result is limited to the frozen eval_core evaluation under BASE, SYSTEM, CONTEXT, and QLoRA. It does not establish results for holdouts, LONG-NEUTRAL, persistence, remediation, or other model families.

license: other applies to this adapter repository. The upstream base model remains governed by its own license.