harshbhx/ratchet-toolcall-train

Ratchet tool-calling: synthetic training data (round 3)

2,940 synthetic training examples for tool calling, written for the third Ratchet round. The task: the model sees 1 to 8 tool definitions (one JSON object per line) and a request, and replies with one JSON object {"name": ..., "arguments": {...}} and nothing else, or {"name": "none", "arguments": {}} when no tool fits. Convention in the training data: an optional argument is included only when the request states its value.

Example row

First row of data/train_v2.jsonl (long strings shortened with ...).

{
  "prompt": [
    {
      "role": "system",
      "content": "You are a tool-calling assistant. Today is 2026-10-24. Given the tools below and a request, reply with one JSON object {\"name\": ..., \"arguments\": {...}} and nothing else. If no tool fits, reply {\"name\": \"none\", \"arguments\": {}}."
    },
    {
      "role": "user",
      "content": "Tools:\n{\"name\":\"search_inbox\",\"description\":\"Search the user's inbox.\",\"parameters\":{\"query\":{\"type\":\"string\"},\"unread_only\":{\"type\":\"boolean\",\"optional\":true}}}\n{\"name\":\"send_payment\",\"description\":\"Send money to a person.\",\"parameters\":{\"recipient\":{\"type\":\"string\"},\"amount\":{\"type\":\"number\"},\"currency\":{\"type\":\"string\",\"enum\":[\"USD\",\"EUR\",\"GBP\",\"INR\"]},\"note\":{\"type\":\"string\",\"optional\":true}}}\n{\"name\":\"cancel_ord ..."
    }
  ],
  "completion": [
    {
      "role": "assistant",
      "content": "{\"name\": \"create_event\", \"arguments\": {\"title\": \"Personal Styling Appointment\", \"start\": \"2026-11-15T11:00:00\", \"duration_minutes\": 45}}"
    }
  ],
  "chat_template_kwargs": {
    "enable_thinking": false
  },
  "provenance": {
    "teacher": "nvidia/Nemotron-3-Ultra-550b-a55b",
    "teacher_license": "OpenMDW-1.1",
    "run_id": "toolsround-2026-10-24-v1",
    "kind": "clear",
    "target_tool": "create_event",
    "task": "tool-calling",
    "labels": "request by teacher; label built in code; agreed by independent teacher answer"
  }
}

Part of Ratchet (code, Apache-2.0). Tool-calling round 3 has three datasets: training data, hard exam (evaluation only) and BFCL exam. The text-to-SQL work is in the text-to-SQL adapter and the tool-calling adapter. Not affiliated with or endorsed by NVIDIA, Nebius, Anthropic or the BFCL authors.

What this is, and is not

• Is: a synthetic training set. Requests were written by NVIDIA Nemotron 3 Ultra (nvidia/Nemotron-3-Ultra-550b-a55b, license OpenMDW-1.1) on Nebius Token Factory. The labels (the correct call) were built in code from the chosen tool and arguments, and an example was kept only where an independent Ultra answer to the same prompt agreed with that label under a strict scorer.
• Is not: customer data, a benchmark, or an exam. Do not use it to report scores; use the exams in the sibling datasets (and do not train on them). No Claude or Anthropic output is in this data.
• Known weakness: the strict scorer cannot agree on free-text arguments, so create_ticket has only 15 rows (target 100) and search_files, send_payment and search_inbox are thin. The set is not balanced across the 20 tools.
• Two runs: one generation run (toolsround-2026-10-24-v1, 2,718 rows) plus a top-up run (toolsround-2026-10-24-topup, 222 rows), both on 24 Oct 2026.

Files

File	What
data/train_v2.jsonl	The 2,940 rows. SHA-256 07cd024aa3f365ba136f2752f32a155b30616cd34a366f2aa419b10ee20db5aa. Its first 2,718 rows are exactly round 3's first file (train.jsonl, SHA-256 b1f1bcf647f460b6c895b7d7518d8a4bb16102bfe370c5cc058509c08d6ab648); the last 222 are the top-up.
tools_catalog.json	The 20 training tools (name, area, description, parameters). Each prompt shows 6 to 8 of them (about 30% of prompts show 1 to 5).
provenance/	report.json and topup_report.json (counts, drops, cost), the generator scripts, the data READMEs
LICENSE.OpenMDW-1.1.txt	License of the model that wrote the data, for reference
MANIFEST.json	SHA-256 of every file

Each row: prompt (a system message and a user message: Tools: then one JSON definition per line, then Request:), completion (the assistant JSON), chat_template_kwargs ({"enable_thinking": false}) and provenance (teacher, teacher license, run id, kind, target_tool, how the label was made).

Kinds: careful 584, clear 1,788, none 568. none rows are near-misses (the wanted tool is not listed) or unrelated requests. 20 distinct target tools plus none.

How it was made

1. Code picks the target tool, kind, register, scene, which tools are shown (the target is always shown except for none), which optional arguments the request must state, and a date phrase from a fixed table (dates are computed from 2026-10-24 in code, never by the model).
2. Ultra writes the request text. Code builds the label.
3. Checks: validator (tool exists, types, enums, emails, airport codes, dates, the optional-argument rule), ids and emails must appear in the request, near-duplicate drop, then the agreement filter (Ultra answers the prompt alone at temperature 0 with thinking off; the strict scorer must equal the label). 3,875 raw items, 3,739 after validation, 2,718 after agreement; the top-up added 222.
4. Overlap check: no training request equals, or is a near-duplicate (token Jaccard above 0.8) of, any request in the hard exam or the BFCL exam (checked: 0 of 150 and 0 of 300).

Generation cost at Token Factory list prices: about $4.28 in all ($3.60 for the first file, $0.68 for the top-up).

License

Apache-2.0. The requests are Ultra outputs and the labels are built in code. OpenMDW-1.1 grants rights to deal in the model materials without restriction and, as read here, puts no restriction on what the model outputs; so we can license the data. See LICENSE.OpenMDW-1.1.txt. The adapter trained on this data is a derivative of NVIDIA's base model and carries NVIDIA's license; that is on its own model card, not here.

Honesty notes

• Adapters in this project were trained on a borrowed GPU. Nebius Token Factory supplied only Nemotron 3 Ultra's answers.
• The training data are synthetic. Nothing here is customer data.
• The exams were frozen (hashed) before any model answered them.
• The tool-calling adapter's result on the hard exam is 136/150 = 90.7% (Wilson 95%: 84.9 to 94.4), above the untrained model and about the same as the teacher; on the BFCL exam it is 241/300 = 80.3% (Wilson 95%: 75.5 to 84.4), about the same as the untrained model. Strict scoring, trained on a borrowed GPU. See the adapter: https://huggingface.co/harshbhx/ratchet-nemotron-nano-4b-toolcall.

Citation

@misc{ratchet2026,
  author = {Harsh Bhanushali},
  title  = {Ratchet: license-clean distillation on open models},
  year   = {2026},
  url    = {https://github.com/harshbhx/ratchet}
}