菜单
Supervised fine-tune a causal LM with TRL `SFTTrainer`. Triggered when the user wants to fine-tune / SFT / instruct-tune / chat-tune a model on conversational, prompt-completion, or text-formatted data. Enforces the literature-first → audit → smoke-test → scale workflow.
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
Set up training and inference on AMD Ryzen AI Max+ 395 / Strix Halo (gfx1151, RDNA 3.5) with TheRock nightly ROCm wheels. Triggered when the host has gfx1151, when `rocminfo` shows Strix Halo, or when the user mentions Strix Halo / Ryzen AI Max / gfx1151 / 128GB unified memory.
Recreate the ml-intern-kit Python environment on a new machine (laptop, rented GPU box, Docker, fresh checkout). Triggered when the user is on a new host, sees ImportError on a core dep (torch/transformers/trl/peft/accelerate), or wants to install flash-attn / unsloth / bitsandbytes after the fact.
Evaluate a trained or downloaded language model with `lm-eval-harness` standard tasks (arc, hellaswag, gsm8k, mmlu, truthfulqa, ifeval, ...). Triggered when the user wants to benchmark, eval, or compare a model — pre- or post-training.
Audit a Hugging Face dataset before training to confirm splits, columns, format, sample rows, distributions, and duplicates. Triggered before any training/fine-tuning script runs, when a user mentions a new dataset, or when you hit a KeyError / format mismatch in a training job.
Submit a training/inference script to Hugging Face Jobs (`hf jobs run`). Triggered when the user wants to run training in the cloud, scale beyond local hardware, or kick off a multi-hour fine-tune. Enforces pre-flight: hub_model_id required, ≥2h timeout for training, single-job validation before batch.
Run a literature-first crawl before writing ANY ML training/fine-tuning/inference code. Spawns an Explore sub-agent that mines papers, citation graphs, methodology sections, and matched HF datasets to produce a ranked list of training recipes attributed to specific published results. Triggered when the user asks to fine-tune, train, or improve a model, or when the user names a task/benchmark and you need a recipe before coding.
| name | train-sft |
| description | Supervised fine-tune a causal LM with TRL `SFTTrainer`. Triggered when the user wants to fine-tune / SFT / instruct-tune / chat-tune a model on conversational, prompt-completion, or text-formatted data. Enforces the literature-first → audit → smoke-test → scale workflow. |
Use the working template at scripts/train_sft.py. Do not write the
trainer from scratch — TRL renames things every release.
research-recipe if you don't already have a paper-
backed recipe. Capture: dataset, method, hyperparameters, reference URL.python scripts/inspect_dataset.py <dataset>. Confirm:
messages, text, or prompt+completion).max_seq_length.configs/sft_default.yaml to configs/<run-name>.yaml
and override model.name_or_path, dataset.name, hyperparameters, and
(when ready to ship) train.push_to_hub: true + train.hub_model_id.python scripts/train_sft.py --config configs/<run-name>.yaml \
--max-steps 20 --max-samples 256
push_to_hub is on, the smoke test creates the
repo — that's fine.- Reference implementation: <paper / GH URL>
- Dataset format verified : <yes — columns: …>
- Model verified : <repo_id, arch, tokenizer>
- push_to_hub : True, hub_model_id=<…>
- Timeout : <≥2h>
- Monitoring : Trackio dashboard URL
accelerate launch scripts/train_sft.py --config configs/<run>.yamlpython scripts/launch_hf_job.py --script scripts/train_sft.py --config configs/<run>.yaml --hardware <flavor> --timeout <≥2h> --hub-model-id <user/repo>In every TrainingArguments / SFTConfig:
disable_tqdm=True
logging_strategy="steps"
logging_first_step=True
logging_steps=10
report_to=["trackio"] # or ["wandb"] if WANDB_API_KEY is set
Loss prints as plain text lines you can grep.
Apply in this order, never changing the user's request:
per_device_train_batch_size, double gradient_accumulation_steps.gradient_checkpointing=True.optim="adamw_8bit" (bitsandbytes).bf16=True (Ampere+).Do not silently switch full SFT to LoRA, drop max_seq_length, or remove
monitoring. Those change what the user gets.