LoRA adapter for Qwen2.5-VL-7B-Instruct, trained with the placebo-controlled objective at α = 0.5, seed 1, with checkpoints from step 50 to step 300 (training dynamics).
Released with the paper Accuracy Is Not Evidence: Measuring and Training Evidence Dependence in Video Temporal Grounding (ICLR 2027 submission, under review). Code, data identifiers and evaluation scripts are in iclropen/evidence-audit-code.
It appears in Figure 9 of the paper.
The root folder is step 300, and checkpoints/step-NNN/ hold the run after NNN of its 300 steps (8 pairs per step). Figure 9 plots these checkpoints.
| Folder | Contents | transformers |
|---|---|---|
| root | final adapter, step 300 | 5.x (tested 5.12.1, peft 0.19.1 or later) |
checkpoints/step-050/ |
checkpoint at step 50 | 5.x (tested 5.12.1, peft 0.19.1 or later) |
checkpoints/step-100/ |
checkpoint at step 100 | 5.x (tested 5.12.1, peft 0.19.1 or later) |
checkpoints/step-150/ |
checkpoint at step 150 | 5.x (tested 5.12.1, peft 0.19.1 or later) |
checkpoints/step-200/ |
checkpoint at step 200 | 5.x (tested 5.12.1, peft 0.19.1 or later) |
checkpoints/step-250/ |
checkpoint at step 250 | 5.x (tested 5.12.1, peft 0.19.1 or later) |
Adapters saved under transformers 4.51 and 5 use different module names. Loading one under the other version raises no error but leaves the LoRA weights unused, so use the version given here. tools/check_lora_keys.py in the code repository checks that an adapter actually loaded.
# pip install "transformers>=5.12" "peft>=0.19.1" qwen-vl-utils
import torch
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from peft import PeftModel
base = "Qwen/Qwen2.5-VL-7B-Instruct"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(base)
model = PeftModel.from_pretrained(model, "iclropen/evidence-audit-qwen7b-pc-a05-s1-dynamics")
The evaluation and audit scripts of the code repository take --model <base> --adapter <this repository or folder>.
The placebo-controlled objective of Section 4 of the paper. It keeps the answer likelihood, weighted by α, and adds two margins for each edit operator: the annotated answer must be more likely on the clean video, and on a video with a matched placebo edit outside the event, than on the video with the event edited. A hinge keeps the drop under the placebo edit within a bound set by the reference model.
| Setting | Value |
|---|---|
| Operator weights | Group-DRO |
| α | 0.5 |
| β | 0.1 |
| Placebo bound | placebo drop ≤ 1.25 × reference drop + 0.10 |
| Training pairs | 2,400 from the Charades-STA training split |
| Optimizer steps | 300 |
| Gradient accumulation | 8 |
| Learning rate | 5e-6 |
| Seed | 1 |
| LoRA | r=16, alpha=32, dropout 0.05, on q/k/v/o and gate/up/down projections |
| Edit operators | swap, mask, blur, tshift, shuffle (two per example) |
| Video input | 2 fps, at most 768 frames |
| Prompt template | timer1_default |
| Trainer | code/27_papo_train_qcdro_v3seln_r5.py |
Anonymous. Accuracy Is Not Evidence: Measuring and Training Evidence Dependence in Video Temporal Grounding. ICLR 2027 submission, under review.
adapter_model.safetensors 190 MBcheckpoints/step-050/adapter_model.safetensors 190 MBcheckpoints/step-100/adapter_model.safetensors 190 MBcheckpoints/step-150/adapter_model.safetensors 190 MBcheckpoints/step-200/adapter_model.safetensors 190 MBcheckpoints/step-250/adapter_model.safetensors 190 MBQwen/Qwen2.5-VL-7B-Instruct