菜单
This skill downloads training results (model weights, training logs, metrics) from a vast.ai GPU instance to the local project. It should be used after remote training completes and BEFORE destroying the instance. Triggers include downloading results, fetching trained model, or collecting training artifacts from vast.ai.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Get competition metadata, data files, and rules using Kaggle MCP. Use when looking up competition details, searching competitions, or fetching competition data file info.
Search, download, and manage Kaggle datasets using Kaggle MCP. Use when finding datasets, checking file structures, or downloading data.
Browse and search Kaggle discussion forums using Kaggle MCP. Use when looking for tips, solutions, or community insights about competitions or datasets.
View and analyze competition leaderboard and submissions using Kaggle MCP. Use when checking rankings, scores, or submission history.
Search, view, save, and run Kaggle notebooks using Kaggle MCP. Use when exploring public notebooks, pushing code to Kaggle, or managing notebook sessions.
Submit to a Kaggle competition (code/notebook submission or file upload). Use when submitting predictions or notebooks to a competition.
| name | vast-gpu-download |
| description | This skill downloads training results (model weights, training logs, metrics) from a vast.ai GPU instance to the local project. It should be used after remote training completes and BEFORE destroying the instance. Triggers include downloading results, fetching trained model, or collecting training artifacts from vast.ai. |
Download all training results from a vast.ai GPU instance to the local outputs/ directory. This skill ensures no artifacts are lost before instance destruction.
NEVER destroy a vast.ai instance before running this download procedure. Always download first, verify files locally, then destroy.
| Category | Files | Destination |
|---|---|---|
| Best model | best_model/ (safetensors, config, tokenizer) | outputs/<exp>/best_model/ |
| Training logs | train_*.txt | outputs/<exp>/ |
| Main log | train_*.log (from /workspace/) | outputs/<exp>/ |
| Metrics | train_metrics.json, timing.json | outputs/<exp>/ |
| Config | config_resolved.json | outputs/<exp>/ |
| Checkpoints | checkpoint-*/ (excluded by default) | outputs/<exp>/checkpoint-*/ |
Before downloading, confirm training has finished:
ssh -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null -o LogLevel=ERROR \
-p <SSH_PORT> root@<SSH_HOST> \
'ps aux | grep train.py | grep -v grep | wc -l'
If the process count is 0, training has completed. Also check for errors:
ssh -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null -o LogLevel=ERROR \
-p <SSH_PORT> root@<SSH_HOST> \
'tail -5 /workspace/train_*.log | tr -cd "[:print:]\n"'
Run the bundled download script:
# Standard download (model + logs + metrics, no checkpoints)
scripts/download_from_vast.sh <SSH_HOST> <SSH_PORT> <EXP_NAME> <PROJECT_ROOT>
# Include checkpoints
scripts/download_from_vast.sh --with-checkpoints <SSH_HOST> <SSH_PORT> <EXP_NAME> <PROJECT_ROOT>
# Preview what will be downloaded
scripts/download_from_vast.sh --dry-run <SSH_HOST> <SSH_PORT> <EXP_NAME> <PROJECT_ROOT>
Example:
scripts/download_from_vast.sh ssh3.vast.ai 24836 exp032_byt5_base_adafactor /path/to/project
After download, verify the local artifacts:
ls -lh outputs/<EXP_NAME>/best_model/
ls outputs/<EXP_NAME>/train_*.txt
cat outputs/<EXP_NAME>/train_metrics.json
Verify model.safetensors is not truncated (should be ~2GB+ for byt5-base).
Only after verifying the download, destroy the instance:
vastai destroy instance <INSTANCE_ID>
scripts/download_from_vast.sh - Rsync-based result download script