LoRA adapter for Qwen2.5-VL-7B-Instruct, trained with GRPO with a temporal-IoU reward on 16 frames per video, checkpoint at step 700. This is the IoU-GRPO model of the main tables.
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 Tables 1, 6–10, 13, 18 and 19 and in Figures 1, 2 and 5 of the paper. The revised version adds Tables 23, 24, 26 and 28 (Appendix F).
Load it with transformers 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-iougrpo-16frames-step700-s0")
The evaluation and audit scripts of the code repository take --model <base> --adapter <this repository or folder>.
GRPO (TRL) with format and temporal-IoU rewards, on 16 frames per video. Its trainer is not part of the release.
| Setting | Value |
|---|---|
| Video input | 16 frames at 224 px |
| Optimizer steps | 700 (this checkpoint is the last step) |
| Questions per step / samples per question | 4 / 8 |
| Reward aggregation | sum, then normalise; no reward scaling |
| Gradient accumulation | 4 |
| Seed | 0 |
| LoRA | r=64, alpha=128, on q/k/v/o and gate/up/down projections |
| Training data | Charades-STA training split |
Anonymous. Accuracy Is Not Evidence: Measuring and Training Evidence Dependence in Video Temporal Grounding. ICLR 2027 submission, under review.
Qwen/Qwen2.5-VL-7B-Instruct