Training
- Base revision:
6e6556e8ce728c7b3e438d75ebf04ec93403dc19
- 174,618 retained samples from the ten-source AURA training mixture; complete examples longer than 15,000 tokens were excluded.
- Full language-model SFT; BF16, FlashAttention 2 and DeepSpeed ZeRO-3.
- 32 GPUs; per-device batch 1, gradient accumulation 2, global batch 64.
- One epoch, 2,729 optimizer steps; learning rate 1e-5, cosine schedule, warmup ratio 0.1, seed 42.
- Video sampling: 1 FPS, maximum 30 frames per video; image/video pixel budget 301,056.
qwen2_vl template with an empty default system message; source-provided system messages were preserved.
- For samples containing silent responses, AURA retains silent answer spans and the final non-silent answer span. Retained silent tokens receive weight 1 / their count in the micro-batch. Samples without silent tokens use ordinary CE. No focal loss was used.
Training finished on 2026-11-08. Final-step training loss: 0.5704. Mean of the final 100 steps: 0.581653. Full-run mean training loss: 0.634191. These are training metrics, not held-out evaluation scores; final evaluation has not been run. See training_summary.json and training_loss.png.
Loading
Tested with Transformers 4.57.1 and PyTorch 2.6.0. The standard Qwen2.5-VL architecture is used; remote model code is not required.
from transformers import AutoModelForImageTextToText, AutoProcessor repo = "leobo09263013/qwen25"processor = AutoProcessor.from_pretrained(repo)model = AutoModelForImageTextToText.from_pretrained( repo, torch_dtype="auto", device_map="auto")
Use the Qwen2.5-VL processor and chat template to prepare image/video inputs. <|silent|> is an intentional assistant response token and can be handled explicitly by downstream applications.
aura_loss.py records the custom training objective; standard Transformers inference does not require it. To continue training with the same objective, explicitly attach the AURA CE hook rather than assuming the metadata in config.json activates a custom loss. The included hook handles the Transformers 4.57 Qwen2.5 composite-config type forwarding behavior on checkpoint reload.
Limitations
Performance and silent-response behavior require downstream evaluation. This model inherits the capabilities and limitations of Qwen2.5-VL-7B-Instruct. Training optimizer states and intermediate checkpoints are not included in this inference model repository.