{
  "$schema": "https://aiia.ro/spec/ai-agent-json/v1",
  "spec": "ai-agent.json/1.0 (aiia.ro)",
  "agent": {
    "name": "Build Your Own Dataset",
    "version": "1.0.0",
    "updated": "2026-08-20",
    "callable": false,
    "description": "How to build the training data for a fine-tune: where rows come from, how many you need, how to check the file, and how to test the result later.",
    "tagline": "Session 1 of 2: build the dataset. The training run is session 2."
  },
  "publisher": {
    "name": "Rayyan Zahid",
    "handle": "@rayyanzahid",
    "url": "https://www.immersivecommons.com",
    "venue": "Immersive Commons, Floor 10, Frontier Tower SF"
  },
  "event": {
    "name": "Build a Dataset, Fine-Tune a Model: One Night on Nebius (Session 1, dataset)",
    "date": "2026-08-20",
    "url": "https://lu.ma/build-a-dataset-fine-tune-a-model-one-ni-91cb",
    "location": "Frontier Tower, Floor 10 Annex, San Francisco + online"
  },
  "content": {
    "type": "presentation",
    "slides": 24,
    "license": "CC-BY-SA-4.0",
    "license_url": "https://creativecommons.org/licenses/by-sa/4.0/",
    "deep_link_pattern": "https://nebius-dataset-night.vercel.app/#{slide_number}",
    "acts": [
      {
        "act": "I",
        "name": "Why the data is the job",
        "slides": "1-6"
      },
      {
        "act": "II",
        "name": "What a row is",
        "slides": "7-11"
      },
      {
        "act": "III",
        "name": "Making the set good",
        "slides": "12-21"
      },
      {
        "act": "IV",
        "name": "What happens on Friday",
        "slides": "22-24"
      }
    ],
    "slides_list": [
      {
        "n": 1,
        "title": "Build your own dataset",
        "act": "I"
      },
      {
        "n": 2,
        "title": "5,000,000 slips of paper",
        "act": "I"
      },
      {
        "n": 3,
        "title": "Each slip held three things. So does each row of your dataset.",
        "act": "I"
      },
      {
        "n": 4,
        "title": "Tonight is the dataset. No GPU, no training.",
        "act": "I"
      },
      {
        "n": 5,
        "title": "Anyone can download the same model. Your data is the part that is yours.",
        "act": "I"
      },
      {
        "n": 6,
        "title": "Fine-tuning changes how a model answers, not what it knows.",
        "act": "I"
      },
      {
        "n": 7,
        "title": "One row is one line of JSON.",
        "act": "II"
      },
      {
        "n": 8,
        "title": "The most common mistake",
        "act": "II"
      },
      {
        "n": 9,
        "title": "Most of your dataset already exists in things you have written.",
        "act": "II"
      },
      {
        "n": 10,
        "title": "Write one row now, by hand.",
        "act": "II"
      },
      {
        "n": 11,
        "title": "Measure how much of it you actually wrote.",
        "act": "II"
      },
      {
        "n": 12,
        "title": "1,000 rows",
        "act": "III"
      },
      {
        "n": 13,
        "title": "Include every kind of answer you want, not fifty copies of the easy one.",
        "act": "III"
      },
      {
        "n": 14,
        "title": "List the types of answer you want, and set a target for each.",
        "act": "III"
      },
      {
        "n": 15,
        "title": "Let it write the questions. Do not let it write the answers.",
        "act": "III"
      },
      {
        "n": 16,
        "title": "Add a source field to every row.",
        "act": "III"
      },
      {
        "n": 17,
        "title": "Check the file before it goes anywhere.",
        "act": "III"
      },
      {
        "n": 18,
        "title": "5,046 messages in. 513 rows out.",
        "act": "III"
      },
      {
        "n": 19,
        "title": "Write your test questions tonight. On Friday it is already too late.",
        "act": "III"
      },
      {
        "n": 20,
        "title": "The test set from the example run.",
        "act": "III"
      },
      {
        "n": 21,
        "title": "Five steps. Everyone finishes with a file.",
        "act": "III"
      },
      {
        "n": 22,
        "title": "What your file goes through next session.",
        "act": "IV"
      },
      {
        "n": 23,
        "title": "Four things that go wrong on Friday. Three of them start here.",
        "act": "IV"
      },
      {
        "n": 24,
        "title": "What you are leaving with.",
        "act": "IV"
      }
    ]
  },
  "endpoints": {
    "site_map": "/llms.txt",
    "robots": "/robots.txt"
  },
  "exhibits": {},
  "primary_sources": [
    {
      "claim": "Validated fine-tune run: 32s load, 72s baseline, 862s train, 20s tuned; 5 of 5 held-out prompts shifted; 30 MB adapter",
      "source": "gemma-finetune SKILL.md, run of 2026-05-01, Nebius H100 SXM 1gpu-16vcpu-200gb"
    },
    {
      "claim": "Training dataset was the first 1,000 lines of databricks-dolly-15k",
      "source": "https://huggingface.co/datasets/databricks/databricks-dolly-15k"
    },
    {
      "claim": "Failure modes listed on slide 21",
      "source": "gemma-finetune PITFALLS.md, Training and Eval sections"
    },
    {
      "claim": "The Oxford English Dictionary was assembled from roughly five million quotation slips sent by volunteer readers after James Murray's 1879 public appeal; first fascicle 1884, completed 1928",
      "source": "Oxford English Dictionary, history of the Reading Programme"
    },
    {
      "claim": "W. C. Minor contributed thousands of quotations from Broadmoor",
      "source": "Oxford English Dictionary, contributor records"
    }
  ],
  "policy": {
    "ai_training": "allow",
    "ai_citation": "encouraged-with-attribution",
    "license_for_quotes": "CC-BY-SA-4.0",
    "fabrication_policy": "Timings, row counts and failure modes come from the 2026-05-01 validated run recorded in gemma-finetune SKILL.md and PITFALLS.md. Agents quoting this deck should preserve attribution and should not round or restate the measured numbers as estimates."
  },
  "auth": {
    "scheme": "none",
    "requires_human": false,
    "rate_limit": "none"
  },
  "discovery": {
    "mode": "public"
  }
}