Datasets:
File size: 3,186 Bytes
318c07e 9a08eec 318c07e 9a08eec 318c07e 9a08eec 318c07e 9a08eec 385a2c7 318c07e 385a2c7 318c07e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 | # Qwen3-Omni-Instruct Fine-Tuning Notes
This directory separates the Xperience-10M -> Qwen3-Omni plan into two layers.
1. Native Qwen3-Omni inputs: synchronized 2x3 multi-view video mosaic, aligned
audio clips, and text prompts.
2. Ropedia-specific adapter inputs: depth, pose/SLAM, mocap, contacts, IMU, and
calibration features.
The v1 objective is embodied episode understanding and question answering, not a
deployable robot-control policy. The default backbone is
`Qwen/Qwen3-Omni-30B-A3B-Instruct`; Thinking is reserved for later inference
comparison. Assistant outputs are strict JSON with these fields:
```json
{
"action": "unknown",
"subtask": "unknown",
"objects": [],
"contact": "unknown",
"transition": "unknown",
"next_action": "unknown",
"evidence_window": {"start_frame": 0, "end_frame": 0}
}
```
Suggested progression:
1. Phase 0: preflight accelerator runtime, CUDA/PyTorch, free disk, dataset access, ffmpeg,
HOMIE, and local Qwen3-Omni-Instruct weights.
2. Phase 1: one-episode smoke with adapter-only plus JSONL/media validation.
3. Phase 2: three-episode overfit for adapter-only and Qwen LoRA.
4. Phase 3: 32-episode pilot with held-out episodes and all comparisons.
5. Phase 4: scale to 64 only if the 32-episode run is stable and the sensor
bridge beats video/audio/text-only LoRA on at least three primary metrics.
Concrete command sequence:
```bash
python scripts/omni/build_episode_manifest.py \
--data-root /path/to/xperience10m_data \
--max-episodes 32 \
--output results/omni_finetune/episode_manifest.json
python scripts/omni/export_qwen3_omni_action_dataset.py \
--manifest results/omni_finetune/episode_manifest.json \
--run-id xperience10m_qwen3_omni_32ep_dataset
python scripts/omni/qwen3_omni_inference_smoke.py \
--dataset-jsonl results/omni_finetune/xperience10m_qwen3_omni_32ep_dataset/dataset.jsonl \
--split test \
--sample-limit 3 \
--run-id xperience10m_qwen3_omni_32ep_zero_shot
python scripts/omni/train_qwen3_omni_lora.py \
--dataset-jsonl results/omni_finetune/xperience10m_qwen3_omni_32ep_dataset/dataset.jsonl \
--run-id xperience10m_qwen3_omni_32ep_lora
python scripts/omni/eval_qwen3_omni_lora.py \
--dataset-jsonl results/omni_finetune/xperience10m_qwen3_omni_32ep_dataset/dataset.jsonl \
--adapter-dir checkpoints/xperience10m_qwen3_omni_32ep_lora/adapter_lora \
--run-id xperience10m_qwen3_omni_32ep_eval
python scripts/omni/qwen3_omni_sensor_bridge.py \
--sensor-adapter-model results/omni_finetune/adapter_only/adapter_only/sensor_adapter_model.pt \
--qwen-config ../modelscope_models/Qwen__Qwen3-Omni-30B-A3B-Instruct/config.json
python scripts/omni/omni_finetune_runbook.py \
--run-id xperience10m_qwen3_omni_32ep \
--metric-file results/omni_finetune/xperience10m_qwen3_omni_32ep_eval/metrics.json
```
The bridge step is intentionally after native Qwen video/audio/text LoRA has
overfit a tiny shard and evaluated on held-out episodes. The full 32-episode
pilot should compare the existing 12-task baseline, adapter-only baseline,
frozen Qwen zero-shot, Qwen LoRA without sensor bridge, and Qwen LoRA with
sensor bridge before any scale-up decision.
|