LoRA adapter for Qwen2.5-VL-7B-Instruct, trained with GRPO with a temporal-IoU reward at 2 fps on our 2,400 training pairs, seed 1, with its checkpoint at step 300.
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.
Trained for the revised version of the paper, where it appears in Table 23 (Appendix F).
checkpoints/step-300/ is the same run after 300 of its 600 steps (1,200 questions), the step-300 row of Table 23.
| Folder | Contents | transformers |
|---|---|---|
| root | final adapter | 4.51 (tested 4.51.3, peft 0.20) |
checkpoints/step-300/ |
checkpoint at step 300 | 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.
# pip install "transformers==4.51.3" "peft==0.20.0" 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-2fps-s1")
The evaluation and audit scripts of the code repository take --model <base> --adapter <this repository or folder>.
GRPO with a temporal-IoU reward, at 2 fps on the same 2,400 training pairs as the placebo-controlled models.
| Setting | Value |
|---|---|
| KL coefficient | 0.04 |
| Samples per question | 8 |
| Questions per step | 4 |
| Warm-up steps | 20 |
| Training pairs | 2,400 from the Charades-STA training split |
| Optimizer steps | 600 |
| Learning rate | 1e-5 |
| 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/29_grpo_native_iou.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-300/adapter_model.safetensors 190 MBQwen/Qwen2.5-VL-7B-Instruct