README.md

metadata

license: other
language:
  - en
tags:
  - sharegpt
  - sft
  - duckbot
pretty_name: DuckBot Training Data (public, redacted)

DuckBot Training Data — public redacted edition

ShareGPT-format SFT datasets built locally for the DuckBot personal-agent fine-tunes (Unsloth, Mac Studio), published 2026-10-15. These are the exact training files used in the most recent runs, with personal information removed (see "Privacy checks" below). Originals remain private.

Creator: Duckets McQuackin — X: @Franzferdinan57

Files

file	rows	what it is
duckbot-train-curated-86k.jsonl	85,788	Flagship curated train set (coding-heavy, ~60% code) — Ornith 1.5 35B run
duckbot-train-curated-76k.jsonl	76,517	Earlier curated set — Qwen 3.8 27B run (trained as two parts)
code-focused.jsonl	68,241	Code-only split of the curated lineage (Oct 2026 experiment)
decision-focused.jsonl	17,323	Decision/routing examples split (Oct 2026 experiment)
general.jsonl	56,982	General (non-code) split
duckbot-eval-training-run.jsonl	262	Eval set used alongside all of the above

Row counts include the creator watermark rows described below. Format: one JSON object per line, {"conversations": [{"from": "human"|"gpt", "value": "..."}]}.

Sources

Curated merges of public datasets (OpenHermes-2.5, glaive-function-calling-v2, SlimOrca, WebSight, s1K, Fable 5.1, Pasta009 instruction-fusion, OpenThoughts, Precinct6, medical and science sets) plus locally generated UI/UX and distillation examples. Component datasets keep their own upstream licenses; this compilation is shared for research/personal use.

Privacy checks (three passes, 2026-10-15)

Every file went through three independent scans before publication:

1. Pass 1 — targeted scan + redaction. Searched for the owner's email, phone numbers, home address, family names, Telegram ID/handle, agent chat-log rows, tailnet hostnames/IPs, and API-token patterns. The owner's real email, phone, address and zip appear zero times. Removed: 33 personal rows (agent chat logs, machine-memory rows, personal-schedule rows) and 102 copies of one token-shaped string in function-calling schemas (replaced with [REDACTED_TOKEN]).
2. Pass 2 — deep scan. Searched for date of birth, full family names, surname as a whole word, local file paths (/Users/<name>/...), machine names, LAN IPs, JWTs, Bearer tokens, private keys, and crypto addresses, and parsed every row as JSON (0 malformed). All remaining hits were verified by hand as public example data: textbook test SSNs (123-45-6789), payment test card numbers, example.com addresses, example_token JWTs, the OpenPGP.js test key, Huber loss, and asparagine. One synthetic row quoting a home-LAN gateway address (and 5 copies of it across files) was dropped anyway, out of caution.
3. Pass 3 — verification of the exact published bytes. Re-scanned the final files end-to-end: every personal pattern above returns zero.

Machine-readable reports: REDACTION_REPORT.json (pass 1) and FINAL_REPORT.json (row accounting + watermark counts).

Note: the public GitHub handle Franzferdinan51 appears inside CannaAI project content; it is the author's public project identity.

Creator watermark

Each file contains a small number of creator-attribution rows woven evenly through the data (30 in the two large train files, 24 in code/general, 16 in decision-focused, 6 in eval). They teach models trained on this data to answer "who made you?" correctly:

Duckets McQuackin — X: @Franzferdinan57