cy0307/ropedia-qwen3-omni-lora-smoke

Ropedia Xperience-10M Qwen3-Omni LoRA Diagnostic Pilot

This repository contains the LoRA adapter from the first selected-episode validation-aware Qwen3-Omni diagnostic pilot for the Ropedia Xperience-10M task-suite project. It is a real held-out-episode run, but it is not a strong model-quality result yet. The adapter is useful as a reproducible baseline, an error-analysis starting point, and a concrete bridge from the public sample task suite toward larger Xperience-10M fine-tuning.

Base Model

• Base model: Qwen/Qwen3-Omni-30B-A3B-Instruct
• Adapter method: LoRA
• Rank: 16
• Alpha: 32
• Dropout: 0.05
• Precision: bf16
• Full-parameter fine-tuning: not included

Data Scope

The selected gated-data split used for this diagnostic pilot was:

Split	Selected episodes	Exported windows
Train	96	2,848
Validation	16	512
Test	16 selected, 14 exported	448

The training job used the train split for optimization and the validation split for validation-loss monitoring. The metrics below are computed on held-out test episodes.

Raw Xperience-10M MP4/HDF5/RRD files and Qwen base weights are not included in this repository.

Held-Out Test Metrics

Metric	Value
JSON validity	0.8750
Action macro-F1	0.0027
Subtask accuracy	0.0067
Transition accuracy	0.8504
Next-action accuracy	0.0246
Contact accuracy	0.6451
Object micro-F1	0.2230
Held-out test windows	448
Held-out exported test episodes	14

The JSON-validity target for the next run is at least 0.98. This pilot falls short of that target, so the next pass should improve structured-output constraints, target formatting, and action/subtask error analysis before expanding the scale.

Related Project Links

• Project website: https://chaoyue0307.github.io/ropedia-xperience-10m-task-suite/
• GitHub repository: https://github.com/chaoyue0307/ropedia-xperience-10m-task-suite
• Artifact dataset: https://huggingface.co/datasets/cy0307/ropedia-xperience-10m-task-suite-artifacts
• Baseline model repository: https://huggingface.co/cy0307/ropedia-xperience-10m-task-baselines
• Official gated dataset: https://huggingface.co/datasets/ropedia-ai/xperience-10m
• Public sample dataset: https://huggingface.co/datasets/ropedia-ai/xperience-10m-sample

Intended Use

Use this adapter to inspect the exact diagnostic LoRA checkpoint produced by the project, compare future structured-output and error-analysis passes against this baseline, and study failure modes on the exported held-out predictions. It should not be presented as a production robot policy or a high-quality embodied foundation model.