Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- sharegpt
|
| 7 |
+
- sft
|
| 8 |
+
- duckbot
|
| 9 |
+
pretty_name: DuckBot Training Data (public, redacted)
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# DuckBot Training Data — public redacted edition
|
| 13 |
+
|
| 14 |
+
ShareGPT-format SFT datasets built locally for the DuckBot personal-agent
|
| 15 |
+
fine-tunes (Unsloth, Mac Studio), published 2026-10-15. These are the exact
|
| 16 |
+
training files used in the most recent runs, with personal information
|
| 17 |
+
removed (see "Redaction" below). Originals remain private.
|
| 18 |
+
|
| 19 |
+
## Files
|
| 20 |
+
|
| 21 |
+
| file | rows | what it is |
|
| 22 |
+
|---|---|---|
|
| 23 |
+
| `duckbot-train-curated-86k.jsonl` | 85,759 | Flagship curated train set (coding-heavy, ~60% code) — Ornith 1.5 35B run |
|
| 24 |
+
| `duckbot-train-curated-76k.jsonl` | 76,488 | Earlier curated set — Qwen 3.8 27B run (trained as two parts) |
|
| 25 |
+
| `code-focused.jsonl` | 68,220 | Code-only split of the curated lineage (Oct 2026 experiment) |
|
| 26 |
+
| `decision-focused.jsonl` | 17,307 | Decision/routing examples split (Oct 2026 experiment) |
|
| 27 |
+
| `general.jsonl` | 56,959 | General (non-code) split |
|
| 28 |
+
| `duckbot-eval-training-run.jsonl` | 256 | Eval set used alongside all of the above |
|
| 29 |
+
|
| 30 |
+
Format: one JSON object per line, `{"conversations": [{"from": "human"|"gpt", "value": "..."}]}`.
|
| 31 |
+
|
| 32 |
+
## Sources
|
| 33 |
+
|
| 34 |
+
Curated merges of public datasets (OpenHermes-2.5, glaive-function-calling-v2,
|
| 35 |
+
SlimOrca, WebSight, s1K, Fable 5.1, Pasta009 instruction-fusion, OpenThoughts,
|
| 36 |
+
Precinct6, medical and science sets) plus locally generated UI/UX and
|
| 37 |
+
distillation examples. Component datasets keep their own upstream licenses;
|
| 38 |
+
this compilation is shared for research/personal use.
|
| 39 |
+
|
| 40 |
+
## Redaction
|
| 41 |
+
|
| 42 |
+
Before publication every file was scanned for personal information
|
| 43 |
+
(emails, phone numbers, addresses, names, IPs, API-token patterns).
|
| 44 |
+
Findings: the owner's real email, phone, home address and zip appear
|
| 45 |
+
**zero** times. Removed/scrubbed for the public edition:
|
| 46 |
+
|
| 47 |
+
- 33 rows total dropped across all files: personal agent chat-log rows
|
| 48 |
+
(Telegram/DuckBot gateway logs), machine-memory rows naming the owner's
|
| 49 |
+
tailnet hosts/IPs, and personal-schedule rows naming his home town.
|
| 50 |
+
- 102 instances of one token-shaped string in function-calling tool
|
| 51 |
+
schemas replaced with `[REDACTED_TOKEN]`.
|
| 52 |
+
|
| 53 |
+
Everything else — including public example data such as textbook test
|
| 54 |
+
SSNs (`123-45-6789`), payment test card numbers, and example.com
|
| 55 |
+
addresses — is unchanged from the training originals. A machine-readable
|
| 56 |
+
`REDACTION_REPORT.json` is included in this repo.
|
| 57 |
+
|
| 58 |
+
Note: the public GitHub handle `Franzferdinan51` appears inside CannaAI
|
| 59 |
+
project content; it is the author's public project identity.
|