Hugging Face's logo

cy0307
/
ropedia-xperience-10m-task-baselines

Robotics
PyTorch
xperience10m_task_baseline_suite
embodied-ai
multimodal
xperience-10m
baseline
evaluation
Model card Files
xet
 Community
ropedia-xperience-10m-task-baselines / results /omni_finetune /HF_UPLOAD.md
cy0307's picture
cy0307
Add files using upload-large-folder tool
367d1e0 verified
preview code
|
Raw
History Blame
1.91 kB

Hugging Face Upload (Qwen3-Omni LoRA Adapter)

The verified public result packages intentionally exclude adapter weights. LoRA weights are uploaded separately as a model repository after the eval package and audit pass.

Current intended target:

  • Model repo: cy0307/ropedia-qwen3-omni-lora-128ep
  • Source package: results/omni_finetune/hf_upload_qwen3_128ep_v6_rank64/
  • Package builder: scripts/omni/prepare_qwen3_lora_hf_package.py
  • Upload script: scripts/omni/upload_qwen3_omni_lora_to_hf.py

Prepare the upload directory from the completed adapter and verified summary:

python3 scripts/omni/prepare_qwen3_lora_hf_package.py \
  --adapter-dir checkpoints/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora/adapter_lora \
  --verified-summary results/omni_finetune/verified_public/xperience10m_qwen3_omni_128ep_multiscale_cap96_v6_rank64_lr5e5_full8gpu_lora_eval_test_full/verified_result_summary.json \
  --output-dir results/omni_finetune/hf_upload_qwen3_128ep_v6_rank64 \
  --repo-id cy0307/ropedia-qwen3-omni-lora-128ep

Upload when network and HF_TOKEN are available:

HF_TOKEN=<your_token> python3 scripts/omni/upload_qwen3_omni_lora_to_hf.py \
  --repo-id cy0307/ropedia-qwen3-omni-lora-128ep \
  --source-dir results/omni_finetune/hf_upload_qwen3_128ep_v6_rank64 \
  --message "Upload Xperience-10M 128-episode Qwen3-Omni LoRA adapter"

The prepared model repo includes PEFT adapter files, tokenizer/processor sidecars when present, a generated model card, and upload_manifest.json with file hashes. It must not include raw Xperience-10M MP4/HDF5/RRD files, Qwen base weights, full FSDP checkpoints, or optimizer state.

Older smoke/pilot adapter material may still exist in results/omni_finetune/hf_upload/ and at cy0307/ropedia-qwen3-omni-lora-smoke. Do not use those paths for the final 128-episode full-run publication.