Files
Table with columns: Folder, Contents, transformers| Folder | Contents | transformers |
|---|
| root | final adapter, step 300 | 4.51 (tested 4.51.3, peft 0.20) |
checkpoints/step-050/ | checkpoint at step 50 | 4.51 (tested 4.51.3, peft 0.20) |
checkpoints/step-100/ | checkpoint at step 100 | 4.51 (tested 4.51.3, peft 0.20) |
checkpoints/step-150/ | checkpoint at step 150 | 4.51 (tested 4.51.3, peft 0.20) |
checkpoints/step-200/ | checkpoint at step 200 | 4.51 (tested 4.51.3, peft 0.20) |
checkpoints/step-250/ | checkpoint at step 250 | 4.51 (tested 4.51.3, peft 0.20) |
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.
The transformers 4.51 runs used gradient checkpointing, under which the LoRA weights of the vision tower received no gradient. They are zero, so these adapters change the language model only.
Usage
# pip install "transformers==4.51.3" "peft==0.20.0" qwen-vl-utilsimport torchfrom transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessorfrom 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-sft-s1-dynamics")
The evaluation and audit scripts of the code repository take --model <base> --adapter <this repository or folder>.
Training
Supervised fine-tuning with the same trainer and data as the placebo-controlled models: the answer likelihood only, with the margin terms switched off.
Table with columns: Setting, Value| Setting | Value |
|---|
| 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 |
| Video input | 2 fps, at most 768 frames |
| Prompt template | timer1_default |
| Trainer | code/27_papo_train_qcdro_v3seln.py |
Retrained model compared with the original
Both adapters were audited against Qwen2.5-VL-7B-Instruct on the first 200 Charades-STA test questions, with the same code under transformers 4.51.3. Γ, the paper's measure of evidence dependence, is in nats per token over the questions both models localise (n), with 95% video-clustered bootstrap intervals. R@0.5 is measured on all 3,720 test questions at 2 fps with greedy decoding, paired by question.
Table with columns: Retrained (this repository), Original run | Retrained (this repository) | Original run |
|---|
| Γ, mask | -0.098 [-0.160, -0.037], n=109 | -0.094, n=108 |
| Γ, swap | -0.058 [-0.100, -0.016], n=109 | -0.036, n=108 |
| R@0.5 change against the base | +8.5 points [7.2, 9.8] (59.2% vs 50.7%) | – |
| Answers with a parsable interval | 100.0% | – |
The original-run values are from the step-300 readout of the original run behind Figure 9.
Citation
Anonymous. Accuracy Is Not Evidence: Measuring and Training Evidence Dependence in Video Temporal Grounding. ICLR 2027 submission, under review.