Update to Adaption

{
  "description": "Generate synthetic datasets using Adaption's Invent-a-Dataset. Use when creating post-training datasets from scratch out of natural language descriptions, to add new capabilities to a language model — especially in a zero-data regime with no seed data available.",
  "included_files": [],
  "name": "adaption-invent",
  "skill_md_contents": "---\nname: adaption-invent\ndescription: Generate synthetic datasets using Adaption's Invent-a-Dataset. Use when \n  creating post-training datasets from scratch out of natural language descriptions, \n  to add new capabilities to a language model — especially in a zero-data regime \n  with no seed data available.\n---\n\n# Adaption Invent Skill\n\nGenerate synthetic training datasets using Adaption's Invent-a-Dataset feature.\n\n## When to use\n\n- Generate synthetic training datasets from natural language descriptions\n- Create domain-specific datasets for post-training language models without any\n  seed data to start from\n- Estimate costs before generating data\n\n## Available MCP Tools\n\n| Tool | Description |\n|------|-------------|\n| `list_invent_domains` | List valid domain and qualified subdomain codes for dataset generation |\n| `generate_dataset` | Estimate or launch generation of a new invented dataset (launches consume credits) |\n\nAfter generation starts, use dataset tools to track progress:\n| Tool | Description |\n|------|-------------|\n| `get_dataset_status` | Poll processing status of the generated dataset |\n| `get_dataset` | Get details of the generated dataset |\n\n## Example Workflows\n\n### Generate domain-specific training data\n\nSteps 1 and 2 are optional — a `dataset_prompt` on its own is enough, and the\ntaxonomy is inferred from it. Supply the codes when you want more control over\nwhat ends up in the dataset.\n\n1. Optionally call `list_invent_domains` to explore available domains and subdomains\n2. Optionally choose domain and subdomain codes matching your use case\n3. Call `generate_dataset` with `estimate: true` to preview cost:\n   ```json\n   {\n     \"domains\": [\"technology\"],\n     \"subdomains\": [\"technology.software_applications\"],\n     \"rows\": 1000,\n     \"estimate\": true\n   }\n   ```\n4. Review estimate, then launch with `estimate: false`\n5. Poll `get_dataset_status` with returned dataset ID until complete\n6. Generated dataset is ready for training — an invented dataset is already a\n   type of adaptive dataset, so it does not need adaptation\n\n### Generate with custom specifications\n\n```json\n{\n  \"name\": \"Customer Support QA\",\n  \"training_type\": \"instruction_dataset\",\n  \"domains\": [\"corporate_business\"],\n  \"subdomains\": [\"corporate_business.customer_service\"],\n  \"rows\": 500,\n  \"dataset_prompt\": \"Generate realistic customer questions about technical software product support scenarios\"\n}\n```\n\n### Control the output through `dataset_prompt`\n\n`dataset_prompt` also carries constraints on output format, tone, and structure,\nwhich is how you pin the shape of the generated rows. Both examples below omit\nthe taxonomy entirely and let it be inferred:\n\n```json\n{\n  \"rows\": 1000,\n  \"dataset_prompt\": \"Question-answer pairs focusing on legal advice, factual explanations, and comparative analysis across diverse topics such as copyright law, government types, and liability. Questions should offer four choices; answers are limited to A, B, C, or D with no additional text.\"\n}\n```\n\n```json\n{\n  \"rows\": 500,\n  \"dataset_prompt\": \"Summaries of used car advertisements in Serbian, featuring detailed descriptions of vehicle conditions, service histories, and equipment packages. Each prompt is an ad summary, and each response is a JSON object containing vehicle_description, service_history, and equipment_packages.\"\n}\n```\n\n### Generate with language expansion\n\n`language_expansion` is either a `translate` or a `localize` spec, and `type` is\nrequired. `sample_rate` (0.01-1) is the share of rows expanded. Credits are\nbilled on the post-expansion row count, so this quotes above `rows`.\n\n```json\n{\n  \"domains\": [\"academic_education\"],\n  \"subdomains\": [\"academic_education.stem\"],\n  \"rows\": 1000,\n  \"language_expansion\": {\n    \"type\": \"translate\",\n    \"languages\": [\"es\", \"fr\", \"de\"],\n    \"sample_rate\": 0.2\n  }\n}\n```\n\nTo localize for country/language pairs instead:\n\n```json\n{\n  \"language_expansion\": {\n    \"type\": \"localize\",\n    \"pairs\": [{ \"country\": \"RS\", \"language\": \"sr\" }],\n    \"sample_rate\": 0.2\n  }\n}\n```\n\n## Domains and Subdomains\n\nDomain codes are flat lowercase strings such as `technology`, `medical`, `legal`,\n`code`, `corporate_business`, `academic_education`, and `personal_finance`.\nSubdomain codes are always qualified with their domain as `domain.subdomain`,\nfor example `medical.symptoms_diagnosis` — a bare code can belong to more than\none domain, so an unqualified value is rejected.\n\nCall `list_invent_domains` for the full catalog with the exact codes. A subdomain\nnaming a domain that is not in your `domains` array is rejected rather than\nsilently added.\n\n## Parameters\n\n`rows` is always required, plus at least one of `dataset_prompt`, `domains`, or\n`subdomains`. With both taxonomy fields omitted, `dataset_prompt` is required and\nthe taxonomy is inferred from it; if nothing can be inferred the request is\nrejected rather than generating across every domain. Supplying `domains` or\n`subdomains` disables that inference.\n\n| Parameter | Description |\n|-----------|-------------|\n| `rows` | Required number of rows to generate, subject to your plan's per-launch cap |\n| `dataset_prompt` | Free-text description of the data you want, up to 10,000 characters |\n| `domains` | Optional array of domain codes |\n| `subdomains` | Optional array of qualified `domain.subdomain` codes |\n| `training_type` | Shape of the generated data: `instruction_dataset` (default) or `preference_pairs` |\n| `name` | Optional name for the generated dataset, up to 255 characters |\n| `language_expansion` | Optional `translate` or `localize` expansion config |\n| `estimate` | Set `true` to preview cost without launching |\n| `idempotency_key` | Optional key for safe retries |\n\n`training_type` here is the shape of the DATA, not the training method:\n`instruction_dataset` produces prompt/completion rows, `preference_pairs`\nproduces ranked completion pairs. The `training_type` inside `hyperparams` on a\ntraining run is a different field meaning LoRA versus full fine-tuning.\n\nGenerate with `preference_pairs` when the dataset will feed an alignment training\nrun — see the `adaption-training` skill, where that run needs\n`training_method: \"alignment\"`.\n\n## Tips\n\n- Start with smaller row counts (100-500) to validate quality\n- Use a specific `dataset_prompt` for better results\n- Once an invented dataset covers the target capability, expand it with more\n  domain-specific or general samples via `augment_dataset` to increase size and\n  diversity before training\n- Use `estimate: true` to check credits before launching\n- Launch responses carry a `next_action` naming the tool to call next — follow it\n- Use `idempotency_key` for safe retries if generation fails\n"
}