菜单
Scaffold a new model + dataset + config triple inside an existing curryTrain project. Activate when the user asks to "add a new model to curryTrain", "scaffold a new experiment", "generate model.py / dataset.py / config.yaml for X", or describes an idea they want to start training. Delegates the actual generation to the scaffolder agent.
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
Run a short, reproducible benchmark of one optimizer step (forward + backward + optimizer step over N microbatches) using the project's registered runtime. Activate when the user asks to "benchmark a training step", "measure throughput", "time one optimizer step", or "smoke test the runtime". Wraps run_accumulated_step from curry_train.benchmark.
Diagnose a training failure or stall by inspecting recent logs, loss curves, OOM traces, NaN events, and config. Activate when the user asks "why did my training crash", "loss went to NaN", "OOM during step X", "training is not improving", or "help me debug this run". Delegates to the failure-diagnoser agent.
Lightning Fabric integration recipe — minimal 5-line setup that gives DDP / FSDP / mixed precision / mixed-precision while keeping a raw PyTorch training loop. Activate when the user asks "Lightning Fabric", "torchrun", "DDP setup", "FSDP setup", "mixed precision", or wires up the launch script.
Bootstrap a new curryTrain project by copying the framework Python template into the user's working directory. Use when the user runs /curry-train:init, asks to "start a new training project with curryTrain", "initialize curryTrain", "scaffold a curryTrain project", or wants the framework code copied locally for editing.
Compare two training runs and produce a concise markdown diff covering config, key metrics, loss curves, and grad-norm trajectory. Activate when the user asks to "compare run A and run B", "diff two experiments", "did this change actually help", or "is this run better than the previous one". This is both the implementation of the action and the methodology guide for variance-aware comparison.
Define an abort condition before launching a run, so that a clearly broken or clearly under-performing run stops itself instead of consuming the full compute budget. Activate when the user asks "when should I kill a run", "abort condition", "early stop a bad run", "kill criterion", or before any expensive run.
| name | new-experiment |
| description | Scaffold a new model + dataset + config triple inside an existing curryTrain project. Activate when the user asks to "add a new model to curryTrain", "scaffold a new experiment", "generate model.py / dataset.py / config.yaml for X", or describes an idea they want to start training. Delegates the actual generation to the scaffolder agent. |
Create a four-file model package (config.py, model.py, checkpoint.py, protocol.py) under curry_train/models/<name>/, plus a starter Hydra config and a minimal dataset adapter. Most of the actual file generation is handed off to the scaffolder agent (see agents/scaffolder.md).
<model-name> (required): kebab-case identifier. Becomes the registered name in register_model(...). Reject names with spaces, slashes, or non-ASCII characters.--task (optional): one of lm, cls, mt, cv, snn. Drives the dataset adapter template the agent picks.--from (optional): a HuggingFace model path the user wants to start from (e.g. --from=Qwen/Qwen2.5-1.5B). The scaffolder uses this to derive shapes and pre-fill config.py.Verify the user is inside a curryTrain project. Look for curry_train/__init__.py and curry_train/models.py. If missing, point the user to /curry-train:init.
Validate the model name. Lowercase, alphanumeric, hyphens only. If conflict with an existing curry_train/models/<name>/ directory, halt and ask whether to overwrite.
Delegate to the scaffolder agent. Spawn the scaffolder subagent with:
skills/stage1-scaffolder/SKILL.md.configs/<name>.yaml.Run preflight asserts. After scaffolding, immediately invoke the checks listed in skills/stage1-preflight-asserts/SKILL.md (zero_grad / mode / shape / leakage). Surface failures back to the user before they try to train.
Suggest the next two stages. Print:
skills/stage2-overfit-single-batch (the gold-standard sanity check).hpo-proposer agent territory in Stage 4.--task=snn), the scaffolder uses primitive-lif-neuron and emits a 4D tensor shape contract (B, T, N, D). For non-SNN, shape stays (B, N, D).--from was passed: the scaffolder must fall back to asking the user to provide a local config / weights path. See primitive-hf-bridge skill for the offline fallback procedure.