Update to Adaption

{
  "description": "Launch and monitor AutoScientist training runs using Adaption MCP. Use when fine-tuning models, checking training progress, or getting hyperparameter recommendations.",
  "included_files": [],
  "name": "adaption-training",
  "skill_md_contents": "---\nname: adaption-training\ndescription: Launch and monitor AutoScientist training runs using Adaption MCP. Use when \n  fine-tuning models, checking training progress, or getting hyperparameter recommendations.\n---\n\n# Adaption Training Skill\n\nLaunch and monitor AutoScientist training runs on the Adaption platform.\n\n## When to use\n\n- List available base models for fine-tuning\n- Get recommended hyperparameters for a dataset\n- Launch AutoScientist training runs\n- Monitor training progress and results\n- List past training runs\n\n## Available MCP Tools\n\n| Tool | Description |\n|------|-------------|\n| `list_training_models` | List base models currently available for AutoScientist training |\n| `list_autoscientist_runs` | List AutoScientist runs, optionally filtered to one dataset |\n| `get_autoscientist_run` | Poll one AutoScientist run, including iteration progress and any public failure message |\n| `recommend_autoscientist_hyperparameters` | Resolve a suitable model and hyperparameters for a ready dataset without launching a run |\n| `create_autoscientist_run` | Launch a new asynchronous AutoScientist training run (spends credits) |\n\n## Example Workflows\n\n### Fine-tune a model with AutoScientist\n\n1. Ensure dataset status is `ready` (check with `get_dataset_status`)\n2. Call `list_training_models` to see available base models\n3. Call `recommend_autoscientist_hyperparameters` with:\n   - `dataset_id`: Your ready dataset\n   - `model`: Optional preferred base model\n   - `augmentation_domain_rows` and `augmentation_general_rows`: the same values\n     you plan to launch with, since they count toward the dataset size the\n     recommendation is sized for\n4. Review the returned `model` and `hyperparams`\n5. Call `create_autoscientist_run` with:\n   - `dataset_id`: Your ready dataset\n   - `model`: Optional base model — the platform picks one when omitted\n   - `column_mapping`: Optional, see below — inferred from the dataset when omitted\n   - `hyperparams`: Optional training configuration, or use the recommended one\n6. Poll `get_autoscientist_run` with `experiment_id` until complete\n\n### Train on preference pairs\n\nAn alignment run trains on ranked completion pairs instead of prompt/completion\nrows. The dataset must have been prepared for it:\n\n1. Generate it with `training_type: \"preference_pairs\"` — see the\n   `adaption-invent` skill — or adapt an existing dataset with that setting\n2. Call `create_autoscientist_run` with `training_method: \"alignment\"`\n3. Map `chosen` and `rejected` in `column_mapping` when the columns are not\n   obvious from the dataset\n\nAn alignment run trains in two stages, supervised first and preference second,\nand reports the first stage until the second begins — so it stays `running`\nthrough the handoff rather than reporting the supervised result as final.\n\n### Augment before training\n\nAutoScientist can augment your dataset before training starts. This is useful for\nincreasing dataset size or adding data closer to the target capability, since many\nmodels need larger volumes before showing a reasonable performance lift.\n\nUse `augmentation_domain_rows` for more samples from the domain already present in\nyour dataset, and `augmentation_general_rows` for samples from other domains that\nare not in it.\n\n```json\n{\n  \"dataset_id\": \"...\",\n  \"augmentation_domain_rows\": 1000,\n  \"augmentation_general_rows\": 500\n}\n```\n\n### Check training history\n\n1. Call `list_autoscientist_runs` to see all runs\n2. Filter by dataset: `list_autoscientist_runs` with `dataset_id`\n3. Get details: `get_autoscientist_run` with `experiment_id`\n\n## Run Options\n\n`create_autoscientist_run` takes these alongside `dataset_id`:\n\n| Parameter | Description |\n|-----------|-------------|\n| `model` | Base model name from `list_training_models`. The platform picks one when omitted |\n| `training_method` | `instruction` or `alignment`. Derived from the dataset when omitted |\n| `data_format` | `chat` or `instruction`, defaulting to `chat`. Applies to raw datasets only — for an adapted dataset the platform resolves the encoding and this is ignored |\n| `column_mapping` | Which dataset columns feed training, see below |\n| `hyperparams` | Training configuration, see below |\n| `augmentation_domain_rows` | Extra rows from the dataset's own domain |\n| `augmentation_general_rows` | Extra rows from other domains |\n| `max_iterations` | How many iterations AutoScientist may run, up to 5 |\n| `target_win_rate` | Win rate to stop at, between 0.5 and 1 |\n| `idempotency_key` | Key for safe retries |\n\n## Column Mapping\n\nA training run accepts a different set of roles than dataset adaptation does, and\nunknown keys are dropped silently instead of rejected — so `context`, `chat`, or\n`image` sent here are ignored without an error.\n\n| Role | Use |\n|------|-----|\n| `prompt` | Column holding the instruction text |\n| `completion` | Column holding the target response |\n| `reasoning_trace` | Column holding a reasoning trace, when the dataset has one |\n| `chosen` | Preferred completion, alignment runs only |\n| `rejected` | Rejected completion, alignment runs only |\n\nOmit `column_mapping` entirely and AutoScientist infers it from the dataset.\n\n## Training Configuration\n\n`hyperparams` is validated strictly, so an unknown key fails the whole request.\nEvery field is optional — omit what you do not want to pin and AutoScientist\nfills it in from the recommendation.\n\n```json\n{\n  \"hyperparams\": {\n    \"training_type\": \"lora\",\n    \"n_epochs\": 3,\n    \"learning_rate\": 2e-5,\n    \"batch_size\": \"max\",\n    \"lora_r\": 16,\n    \"lora_alpha\": 32,\n    \"lr_scheduler_type\": \"cosine\",\n    \"warmup_ratio\": 0.1\n  }\n}\n```\n\n| Parameter | Accepted values |\n|-----------|-----------------|\n| `training_type` | `lora` or `full` — LoRA versus full fine-tuning |\n| `n_epochs` | 1-20 |\n| `learning_rate` | 1e-8 to 1e-2 |\n| `batch_size` | `\"max\"` or a positive integer |\n| `lora_r` | 1-64 |\n| `lora_alpha` | Must equal `lora_r` or twice `lora_r` |\n| `lora_dropout` | 0-1 |\n| `lora_trainable_modules` | `\"all-linear\"` or a comma-separated module list |\n| `lr_scheduler_type` | `linear`, `cosine`, or `constant` |\n| `min_lr_ratio` | 0-1 |\n| `scheduler_num_cycles` | 0-4 |\n| `warmup_ratio` | 0-1 |\n| `max_grad_norm` | Non-negative |\n| `weight_decay` | Non-negative |\n| `train_on_inputs` | Boolean |\n| `dpo_beta` | 0.05-0.9, alignment runs only |\n| `dpo_normalize_logratios_by_length` | Boolean, alignment runs only |\n| `rpo_alpha`, `simpo_gamma` | Alignment runs only, and both cannot be positive at once |\n\n## Training Status\n\nThe `get_autoscientist_run` response reports progress as a count of finished\niterations plus the metrics of the best one — there are no per-iteration metrics:\n\n- `status`, `created_at`, and `completed_at`\n- `iterations_completed` against `max_iterations`\n- `best_win_rate` against `target_win_rate`\n- `best_hyperparams` — the configuration that produced the best iteration\n- `download_available` once a successful run has a trained model\n- `error` with the public failure message when the run failed\n\n## Tips\n\n- Use `recommend_autoscientist_hyperparameters` before launching to optimize configuration\n- Training runs consume credits — verify dataset is ready first\n- Poll `get_autoscientist_run` every 30-60 seconds during training\n- Launch responses carry a `next_action` naming the tool to call next — follow it\n- Use `augmentation_domain_rows` for domain-specific enrichment and\n  `augmentation_general_rows` for general or domain-diverse enrichment\n"
}