| { |
| "dataset_version": "2026-07-20-mimo-v11-final-r7", |
| "canonical_manifests": { |
| "sample": { |
| "schema_version": "avrtg.sample.v1", |
| "files": [ |
| "samples_manifest_train.jsonl", |
| "samples_manifest_val.jsonl", |
| "samples_manifest_test.jsonl" |
| ] |
| }, |
| "query": { |
| "schema_version": "avrtg.query.v1", |
| "files": [ |
| "queries_manifest_train.jsonl", |
| "queries_manifest_val.jsonl", |
| "queries_manifest_test.jsonl" |
| ], |
| "model_target": "total_gt = union(audio_gt, visual_gt)", |
| "diagnostic_fields": [ |
| "audio_gt", |
| "visual_gt", |
| "av_overlap_gt" |
| ] |
| } |
| }, |
| "prompt": { |
| "policy_id": "avrtg.temporal_grounding.union.v1", |
| "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>", |
| "thinking_trace": false |
| }, |
| "model_output": { |
| "content_type": "strict_json_inside_answer_tags", |
| "template": "<answer>{\"intervals\":[[start_second,end_second], ...]}</answer>", |
| "top_level_keys": [ |
| "intervals" |
| ] |
| }, |
| "derived_manifests": { |
| "grpo": { |
| "files": [ |
| "swift_rlvr_train.jsonl", |
| "swift_rlvr_val.jsonl", |
| "swift_rlvr_test.jsonl" |
| ], |
| "solution_intervals_source": "query.total_gt", |
| "diagnostic_metadata_not_exposed_in_prompt": true |
| } |
| } |
| } |
|
|
