| # Reproducibility Contract |
|
|
| This file defines what can be reproduced from the public repo and the official |
| Xperience-10M sample, what each command should produce, and which results remain |
| outside the current public data scope. |
|
|
| ## Scope |
|
|
| | Layer | Reproducible now | Current scope | |
| | --- | --- | --- | |
| | Sample download | Yes, from `ropedia-ai/xperience-10m-sample` or ModelScope sample mirror | Sample card lists `cc-by-nc-4.0`; raw data is not redistributed in this repo. | |
| | Minimal baselines | Yes | One public sample episode, chronological split. | |
| | Unified 20-task suite | Yes; the historical provenance bundle requires `annotation.hdf5` plus `h5py` or HOMIE Toolkit for regeneration | Uses the current 8,546-d synchronized multimodal feature contract, the same 20-frame windows, and the same chronological split. | |
| | Neural MLP heads | Yes, when `torch` is installed | Compact task heads only, not a foundation model. | |
| | Website figures and charts | Yes | Generated from committed metrics and sample thumbnails. | |
| | Public bundle contents | Yes | Covers public repo and prepared HF bundles. | |
| | Multi-episode Qwen3-Omni LoRA pilot | Yes, as a public-safe verified result package | The selected 96/16/16 episode split produced verified held-out packages; the latest v6 package records 34,269 exported multiscale windows and 4,032 held-out predictions. Public readers can inspect the package, but rerunning requires gated Xperience data and base-model weights. | |
| | Owner-side staged Qwen3-Omni v6 reproduction | Yes, on the private staged GPU host only | The staged host has the exported media cache, path-rewritten JSONL, Qwen3-Omni base-model cache, v6 adapter, HF mirrors, and a one-sample smoke with `exit_code=0` on 2026-06-14. | |
|
|
| ## Environment |
|
|
| Use Python 3.12 when possible. The current public scripts depend on the HOMIE |
| toolkit environment plus lightweight plotting and Hub tooling. |
|
|
| ```bash |
| git clone https://github.com/Ropedia/HOMIE-toolkit.git |
| python3.12 -m venv .venv |
| source .venv/bin/activate |
| pip install -r HOMIE-toolkit/requirements.txt huggingface_hub hf_xet |
| pip install -r ropedia-xperience-10m-task-suite/requirements.txt |
| pip install torch |
| ``` |
|
|
| ## Data |
|
|
| Download the public sample from Hugging Face: |
|
|
| ```bash |
| hf download ropedia-ai/xperience-10m-sample \ |
| --repo-type dataset \ |
| --local-dir data/sample/xperience-10m-sample |
| ``` |
|
|
| If Hugging Face access is unavailable in your environment, use the included |
| ModelScope helper: |
|
|
| ```bash |
| python scripts/omni/download_sample_modelscope.py \ |
| --output-dir data/sample/xperience-10m-sample \ |
| --mode all-training |
| ``` |
|
|
| `--mode all-training` downloads `annotation.hdf5` and the six MP4 streams while |
| skipping `visualization.rrd`. |
|
|
| The sample card points to HOMIE Toolkit for inspecting videos and annotations. |
| When `visualization.rrd` is downloaded for human inspection, open it with Rerun |
| 0.29.0. The `.rrd` viewer artifact is not used by the training/evaluation |
| scripts and is excluded from public publication bundles. |
|
|
| ## Core Commands |
|
|
| Run these from the repo root after setting `WORKSPACE` to the folder that owns |
| `data/sample/xperience-10m-sample`. |
|
|
| ```bash |
| export WORKSPACE=/path/to/workspace |
| |
| python scripts/train_min_action_model.py --workspace "$WORKSPACE" |
| python scripts/train_all_modalities_model.py --workspace "$WORKSPACE" |
| |
| python scripts/episode_task_suite.py \ |
| --workspace "$WORKSPACE" \ |
| --include-neural |
| |
| python scripts/research_direction_taxonomy.py |
| python scripts/research_direction_extension_tasks.py |
| python scripts/tier2_task_suite.py |
| python scripts/build_unified_task_suite.py |
| python scripts/build_unified_task_model_radar.py |
| python scripts/task_walkthroughs.py |
| python scripts/validate_source_alignment.py |
| python scripts/build_evaluation_protocol.py |
| python scripts/generate_visualizations.py |
| python scripts/render_overview_figures.py |
| python scripts/render_task_suite_infographic.py |
| python scripts/export_modality_atlas_assets.py |
| python scripts/build_brand_assets.py |
| python scripts/build_figure_index.py |
| python scripts/validate_website_integrity.py |
| python scripts/validate_task_surface.py |
| python scripts/validate_scope_claims.py |
| python scripts/build_artifact_index.py |
| python scripts/validate_mirror_parity.py |
| python scripts/validate_publication_package.py |
| ``` |
|
|
| `scripts/tier2_task_suite.py` has a historical file name, but it now regenerates |
| provenance rows inside the unified 20-task suite. It can use HOMIE Toolkit when |
| present, or a direct `h5py` fallback for the public sample's caption JSON. It |
| reads the local raw `annotation.hdf5` only to regenerate interaction/object |
| targets; the raw HDF5 is still ignored by git and excluded from public bundles. |
|
|
| ## Owner-Side Staged Qwen3-Omni v6 Reproduction |
|
|
| This section is for the private staged GPU host, not for public reruns from the |
| GitHub repo alone. It preserves the verified result path after the original |
| training host is released. |
|
|
| Expected private staging layout: |
|
|
| | Item | Staged path | |
| | --- | --- | |
| | Staging root | `/mnt/kgc/chaoyue/ropedia-h20-side` | |
| | Repo | `<staged-repo-root>` | |
| | Qwen3-Omni base model | `/mnt/kgc/chaoyue/ropedia-h20-side/modelscope_models/Qwen__Qwen3-Omni-30B-A3B-Instruct` | |
| | v6 adapter | `checkpoints/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora/adapter_lora` | |
| | Staged eval JSONL | `results/omni_finetune/xperience10m_qwen3_omni_128ep_multiscale_cap96_v5_full8gpu_lora_dataset/dataset_a100_eval.jsonl` | |
| | Private handoff manifest | `/mnt/kgc/chaoyue/ropedia-h20-side/STAGING_MANIFEST_20260614.md` | |
|
|
| The staged JSONL has the same 34,269 rows as the original export JSONL, with |
| exported media paths rewritten from the training-host repo root to the private |
| staging root. Raw upstream Xperience-10M source files are not required for this |
| train/eval cache reproduction and were not copied because the selected raw |
| source tree is about 278 GB. |
|
|
| Run this from the staged repo: |
|
|
| ```bash |
| cd <staged-repo-root> |
| CUDA_VISIBLE_DEVICES=0,1,2,3 \ |
| RUN_ID=a100_repro_qwen_v6_eval_smoke1_manual \ |
| SAMPLE_LIMIT=1 \ |
| MAX_NEW_TOKENS=1 \ |
| scripts/omni/run_private_gpu_qwen3_v6_repro_smoke.sh |
| ``` |
|
|
| The launcher first applies/checks the narrow Transformers Qwen3-Omni |
| video-feature compatibility patch. The expected compatible installed source |
| hash is `da5feea4afc11767db3ca7eedb85ac129c66605643dadc6272c4288b03be7d25`; |
| the known incompatible pre-patch hash is |
| `2aa5752c32965dbaeee230a016afbbbb30d459a46a12c88c1d6f712e12ba95ad`. |
|
|
| Verified staged-GPU smoke evidence from 2026-06-14: |
|
|
| | Field | Value | |
| | --- | --- | |
| | Run id | `a100_repro_qwen_v6_eval_smoke1_preflight_busy_20260614` | |
| | Exit code | `0` | |
| | Samples | `1` | |
| | JSON validity | `1.0` | |
| | Transition accuracy | `1.0` | |
| | Contact accuracy | `1.0` | |
| | Object micro-F1 | `0.28571428571428575` | |
| | Metrics path | `results/omni_finetune/a100_repro_qwen_v6_eval_smoke1_preflight_busy_20260614/metrics.json` | |
|
|
| ## Expected Public Outputs |
|
|
| | Command group | Expected artifacts | |
| | --- | --- | |
| | Minimal baselines | `results/min_action_model/`, `results/min_all_modalities_action_model/`, metrics and model weights | |
| | Unified 20-task suite | `TASK_SUITE_20.md`, `docs/data/task_suite_20.json`, `results/episode_task_suite/summary_report.json`, per-task `metrics.json`, predictions, confusion matrices, and the historical `tier2_task_suite` provenance bundle | |
| | Unified 20-task model radar | `docs/data/unified_task_model_radar.json`, `docs/assets/charts/unified_task_model_radar.svg` | |
| | Neural heads | `results/episode_task_suite/neural_mlp/**/metrics.json`, histories, model checkpoints | |
| | Research directions | `results/episode_task_suite/research_directions/`, `docs/data/research_directions.json` | |
| | Direction probes | `results/episode_task_suite/research_direction_extensions/`, `docs/data/research_direction_extensions.json` | |
| | Walkthroughs | `results/episode_task_suite/task_walkthroughs/`, `docs/data/task_walkthroughs.json` | |
| | Task surface integrity | `docs/data/task_surface_integrity.json` | |
| | Source alignment | `SOURCE_ALIGNMENT_AUDIT.md`, `docs/data/source_alignment_audit.json` | |
| | Evaluation protocol | `EVALUATION_PROTOCOL.md`, `docs/data/evaluation_protocol.json` | |
| | Figures | `docs/assets/*.png`, `docs/assets/charts/*.svg` | |
| | Brand assets | `docs/assets/brand/*.png`, `docs/favicon.png`, `docs/apple-touch-icon.png`, `docs/data/brand_assets.json` | |
| | Figure index | `FIGURE_INDEX.md`, `docs/data/figure_index.json` | |
| | Modality atlas | `docs/data/modality_atlas.json`, `docs/assets/modalities/*` | |
| | Website integrity | `docs/data/website_integrity.json` | |
| | Release reports | `docs/data/artifact_index.json`, `docs/data/mirror_parity.json`, `docs/data/publication_audit.json`, `docs/data/scope_claims_audit.json` | |
|
|
| ## Exact-Match Reproduction Record |
|
|
| The last full metric reproduction run was completed on **2026-05-30 |
| Asia/Singapore** from a fresh output directory outside the repo. It rebuilt the |
| minimal baselines, all-modality baselines, and the original core task artifacts |
| from the local public sample. The regenerated metrics matched the committed |
| artifacts after float normalization; the current public framing now indexes |
| those artifacts together as one 20-task suite. |
|
|
| Evidence: |
|
|
| - [`notes/reproducibility_audit.md`](notes/reproducibility_audit.md) |
| - [`docs/data/reproducibility_matrix.json`](docs/data/reproducibility_matrix.json) |
|
|
| ## Non-Reproducible From This Public Repo Alone |
|
|
| The following require gated data, large model weights, or private compute |
| state, so this repo does not provide public reproduction for: |
|
|
| - rerunning the multi-episode Qwen3-Omni LoRA pilot from raw gated data, |
| - full Xperience-10M-scale pretraining, |
| - raw Xperience-10M video or annotation redistribution, |
| - full Qwen weights or large full checkpoints. |
|
|
| Before interpreting any Qwen3-Omni result, read |
| [`docs/data/scope_claims_audit.json`](docs/data/scope_claims_audit.json), |
| [`results/omni_finetune/DATA_ACCESS_STATUS.md`](results/omni_finetune/DATA_ACCESS_STATUS.md) |
| and |
| [`results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md`](results/omni_finetune/MULTI_EPISODE_ACCESS_STATUS.md). |
|
|
