菜单
Start, stop, pause, resume, and monitor model training runs. Check training progress, view metrics, and get GPU utilization during training. Use when asked to train a model, check training status, view loss curves, stop a run, or see training logs.
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
Read and control I2C and SPI peripherals on Sipeed boards (LicheeRV Nano, MaixCAM, NanoKVM).
Submit development specs, approve execution plans, monitor workers, retry failed tasks, check job status, and manage orchestration jobs. Use when asked to start a development job, check worker progress, approve a plan, retry something, or list orchestrator jobs.
Generate synthetic training data, manage seeds, preview synthesis jobs, and work with the data factory. Use when asked to generate synthetic data, create training examples, augment a dataset, manage seeds, or check factory job status.
Browse, compare, evaluate, and manage trained models. View leaderboard, trends, lineage, and artifacts. Deploy models to Ollama. Use when asked about models, model comparison, which model is best, model evaluation, deploying a model, or downloading a model.
Check BashGym system status, GPU utilization, hardware info, and aggregate statistics. Use when asked about system health, what's running, GPU usage, memory, disk space, or general system stats.
Browse, promote, demote, classify, and generate training examples from Claude Code execution traces. Use when asked about traces, gold traces, trace counts, promoting or demoting traces, generating examples, exporting training data, or syncing traces.
| name | training |
| description | Start, stop, pause, resume, and monitor model training runs. Check training progress, view metrics, and get GPU utilization during training. Use when asked to train a model, check training status, view loss curves, stop a run, or see training logs. |
Manage SFT/DPO/GRPO training runs on local hardware.
scripts/api.sh POST /training/start '{
"strategy": "sft",
"base_model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
"selected_repos": [],
"epochs": 3,
"batch_size": 4,
"learning_rate": 2e-5
}'
Strategy options: sft, dpo, grpo
Empty selected_repos means train on all gold traces.
scripts/api.sh GET /training/{run_id}
Returns: epoch, loss, learning rate, elapsed time, estimated completion.
scripts/api.sh GET /training
scripts/api.sh POST /training/{run_id}/pause '{}'
scripts/api.sh POST /training/{run_id}/resume '{}'
Destructive — confirm with user first.
scripts/api.sh POST /training/{run_id}/stop '{}'
scripts/api.sh GET /system/info
Report GPU utilization %, memory used/total, temperature if available.
User: "Start training on my gold traces" → Call POST /training/start with defaults. Report run_id and initial status.
User: "How's training going?" → Call GET /training to find active run. Call GET /training/{run_id} for metrics. Also call GET /system/info for GPU stats. Report epoch, loss, GPU %.
User: "Stop the current training" → Confirm with user. Find active run_id. Call POST /training/{run_id}/stop.