mrfarazi/tgac-llava15-stage2

S3 (margin loss, LoRA frozen) (stage2)

Checkpoint stage2 from The Gate Always Closes: On Injecting Auxiliary Signals into Frozen Vision-Language Models (ACCV 2026, arXiv:2607.23335). Paper label: S3 (margin loss, LoRA frozen). Code and the reproduction scripts: https://github.com/mrfarazi/the-gate-always-closes.

What this is

Attention-bias design with the Stage-2 margin-loss gates (LoRA frozen); stage2_checkpoint.pt holds beta and the head / layer gates. Stage-1 checkpoint (detach_bias) plus stage2_checkpoint.pt (beta, head/layer gates).

field	value
kind	hra_stage2
variant (VARIANT_CONFIGS key)	-
backbone	llava-hf/llava-1.5-7b-hf (not redistributed; downloaded from its own repo)
learned gate	beta = 0.0000, softplus_beta = 0.6931, stage2_beta = -8.4962
trainer	see variant.json
files	hra_modules.pt, lora, stage2_checkpoint.pt, variant.json

variant.json records the configuration the graph module must be rebuilt with; the release code reads it (hra.checkpoints). The lora/ adapter is a standard PEFT adapter for the base model.

Load

from hra import hub
from hra.checkpoints import load_for_eval
model, processor, wrapper = load_for_eval("hra", "stage2", "cuda")   # downloads this repo

Behavioural knock-out (from the paper's result files)

No knock-out file for this checkpoint in results/knockouts/.

Licence and citation

Code and these adapter weights: MIT. The base model's own licence applies to the base weights (llava-hf/llava-1.5-7b-hf). Please cite the paper:

@InProceedings{Farazi_2026_ACCV,
    author    = {Farazi, Moshiur and Ramasinghe, Sameera and Ciftler, Bekir Sait and Turza, Mahbub Ahmed and Rahman, Shafin},
    title     = {The Gate Always Closes: On Injecting Auxiliary Signals into Frozen Vision-Language Models},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {December},
    year      = {2026}
}