Instructions to use TimeDuet/TimeDuet-Qwen with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use TimeDuet/TimeDuet-Qwen with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Omni-7B") model = PeftModel.from_pretrained(base_model, "TimeDuet/TimeDuet-Qwen") - Notebooks
- Google Colab
- Kaggle
Download evaluation_protocol.json from TimeDuet/TimeDuet-Qwen: direct link, hf CLI and curl.
- Browser
- Download file 2.59 kB
-
https://huggingface.co/TimeDuet/TimeDuet-Qwen/resolve/main/evaluation_protocol.json
- Command line
-
hf download hf://TimeDuet/TimeDuet-Qwen/evaluation_protocol.json
-
curl -L -o evaluation_protocol.json https://huggingface.co/TimeDuet/TimeDuet-Qwen/resolve/main/evaluation_protocol.json
2.59 kB
| { | |
| "policy_id": "avrtg.r7.common.max1024.v1", | |
| "status": "canonical_for_new_reevaluations", | |
| "split": "R7 held-out test", | |
| "expected_queries": 2202, | |
| "data": { | |
| "validation_sha256": "960e2a04502418f44ea23ca0a591609995379f1c636b7a6db80c017c4cf7237a", | |
| "validation_query_instances": 2202 | |
| }, | |
| "prompt": { | |
| "name": "canonical union-localization prompt", | |
| "system": "You are an audio-visual temporal grounding model.", | |
| "user_template": "<video>\nLocalize every time interval where the following event is present\nin the audio stream, the visual stream, or both:\n\n{query}\n\nReturn JSON inside <answer> tags. Use seconds relative to t=0 of the full video.\nReturn all matching intervals in chronological order, merging overlapping or contiguous intervals.\n\nRequired output format:\n<answer>{\"intervals\":[[start_second,end_second], ...]}</answer>" | |
| }, | |
| "output_schema": "<answer>{\"intervals\":[[start,end], ...]}</answer>", | |
| "generation": { | |
| "do_sample": false, | |
| "temperature": 0.0, | |
| "top_p": null, | |
| "top_k": null, | |
| "num_beams": 1, | |
| "max_new_tokens": 1024, | |
| "termination": [ | |
| "model_native_eos", | |
| "official_conversation_separator" | |
| ], | |
| "answer_wrapper_early_stop": false | |
| }, | |
| "audio": { | |
| "scope": "full", | |
| "local_format": "16-kHz mono", | |
| "source": "same verified MP4" | |
| }, | |
| "video": { | |
| "scope": "full timeline", | |
| "sampling_fps": 1.0, | |
| "timestamps_sec": "0,1,2,...", | |
| "experiment_frame_cap": null, | |
| "experiment_pixel_cap": null, | |
| "spatial_preprocessing": "official model-native processor" | |
| }, | |
| "input": { | |
| "video_sampling_fps": 1.0, | |
| "experiment_frame_cap": null, | |
| "experiment_pixel_cap": null, | |
| "full_mp4_audio_required": true, | |
| "use_audio_in_video": true, | |
| "return_audio": false, | |
| "talker_enabled": false, | |
| "batch_size_per_gpu": 1, | |
| "oom_policy": "Fail and record the row; do not silently reduce frames, pixels, audio, or output length." | |
| }, | |
| "execution": { | |
| "queries_per_gpu_at_once": 1, | |
| "sharding": "deterministic and disjoint", | |
| "resume": true, | |
| "required_generation_audit": [ | |
| "generated_tokens", | |
| "termination_reason", | |
| "cap_hit", | |
| "provider_finish_reason_if_available" | |
| ] | |
| }, | |
| "parser": "strict/common canonical parser", | |
| "metric_version": "avrtg.evidence-region.v2", | |
| "failure_policy": "no reinference or manual repair; empty prediction and localization zero", | |
| "integrity_requirements": { | |
| "errors": 0, | |
| "duplicates": 0, | |
| "missing": 0, | |
| "extra": 0, | |
| "frame_mismatch": 0, | |
| "empty_audio": 0 | |
| } | |
| } | |