--- license: cc-by-4.0 language: - en task_categories: - text-generation tags: - task-extraction - structured-output - voice-notes - synthetic - on-device size_categories: - 1K\nTasks:\n` and the answer as `{"tasks":[{"t":"...","c":"..."}]}`. I used both forms of every note, which doubles the rows. ## What is not included, and why - Real notes. I and five friends gave real brain-dump notes for training, validation and testing. They are private and are not published: 325 real notes (140 in training, 59 in validation, 126 in test) plus 10 notes I typed by hand. - A set of 966 notes (781 training, 93 validation, 92 test) that I drafted with ChatGPT's help. I have not approved them for release, so they are left out entirely. - Rows that failed my checks: 9 rows removed as leaks (8 training, 1 test; see check B) and 2 training rows removed because the text looked like an email or a web address (check A). The numbers in my write-up for the sealed test set (617 notes) include the 126 real notes and the 92 withheld notes, so they cannot be reproduced from this repository. The public test split holds the other 398 notes. ## Evaluation files `eval/test_answers.jsonl` has one row per public test note (`id` matches `data/test.jsonl`). For each of five systems it holds the answer (`tasks`) and the judge's verdicts: - `heed_qwen3_0.6b_int8_phone`: the Heed model (Qwen3-0.6B fine-tuned, int8) running on a phone - `claude_sonnet_5.5_key`: the reference labels from `data/test.jsonl`, judged like any other answer. It is the key, so it is not an independent system - `claude_haiku_5.5`, `gemini_3.5_flash`, `gpt_5.4_mini`: asked with the same labelling instructions as the key `judge_gemini_3_1_pro` is the first judge (Gemini 3.1 Pro, blind, temperature 0, run on every note): per task `verdict` (good, fixable, wrong), `category` (right, wrong, either) and `merged`, plus a `missed` list. The verdict list is in the same order as `tasks`. `judge_gpt_5_4` is a second judge (GPT-5.4) on a random subset of 64 public notes (null elsewhere, and null where its reply was not valid JSON). `eval/judge_prompt.txt` is the frozen judge rubric. The judge got that text, then the category list, the plain-text note and the numbered tasks. `eval/score.py` has no dependencies: ``` python3 eval/score.py # first judge, all public notes python3 eval/score.py --judge gpt # second judge, notes it also judged ``` Precision is good / (good + fixable + wrong), recall is good / (good + fixable + missed), F1 is computed on the totals. The script also runs a paired bootstrap (4,000 resamples of the notes, seed 0) for each system minus Heed. Results on the 398 public test notes (first judge): | System | Judged task F1 | Minus Heed [95% interval] | |---|---|---| | Heed (Qwen3-0.6B int8, phone) | 74.2 | | | Claude Sonnet 5.5 (the key) | 94.7 | +20.5 [+18.2, +22.9] | | Gemini 3.5 Flash | 91.5 | +17.4 [+15.1, +19.7] | | Claude Haiku 5.5 | 83.9 | +9.8 [+7.1, +12.5] | | GPT-5.4 mini | 50.5 | -23.7 [-27.0, -20.4] | With the second judge on its 61 public notes judged for all five systems: Heed 73.6, Sonnet key 87.2, Gemini Flash 84.9, Haiku 81.6, GPT-5.4 mini 64.8. The order of the systems is the same, the gap to GPT-5.4 mini is smaller, and the intervals are wide on 61 notes. ## Checks I ran on the public rows Check A, personal information. Every public row (note, task text and category text, 8,377 rows before removals) was scanned with patterns for emails, phone numbers (Indian and international), web addresses, street addresses with numbers, 12-digit Aadhaar-like numbers, PAN-like codes, card-like 13 to 19 digit numbers, IBANs and postcodes. Result: 2 rows removed (one mention of an email provider and one booking-site address); 24 rows kept with a report-only hit (23 street-address-like patterns that are made-up places, clinic names or unit sizes, and 1 five-digit amount). No phone, card, Aadhaar, PAN or IBAN hits. I also built a list of 53 capitalised names that occur only in the real notes and checked the public rows against it: 1,835 rows contain at least one of them. A few of them appear in hundreds of rows, which looks like names the models use a lot for made-up people, so I treated this as a flag and not as a reason to remove a row. Check B, leaks from the real notes. Every public row was compared with all 335 real notes (training, validation and test, including the typed ones) using character 5-gram and word 4-gram overlap and the longest common run of words. A row counts as a leak if at least 25% of its word 4-grams occur in one real note, or the two share a run of 8 or more words. 9 rows were leaks and were removed: 1 `simulated` (the source written with 10 real notes as style examples), 1 `curated`, 1 `targeted`, 5 `long_braindump` in training and 1 `long_braindump` in test. Across the other 8,368 rows the largest word 4-gram containment is below 0.25 and the 99th percentile is 0.06. After removals there are no identical notes inside or across splits. ## Limitations - One labeller family: the labels carry Claude Sonnet's habits (how it splits and words tasks, which category it picks when two fit). A model trained on them learns that style. The judge is a model too. - English only, with many Indian, Irish, Australian, South African and other English-speaking voices, but all written, not spoken. Real speech-to-text adds word errors that these notes do not have. - The notes lean on a few templates of life (family, work, shopping), and almost no `watch` or `ideas` items appear in the `targeted` source. - 4% of rows have an empty list on purpose, which a model can learn to over-use. - The test split is synthetic. It says little about how a model does on notes from real people. ## Licence CC-BY-4.0. The notes and labels were produced with Claude, Gemini and GPT models. Check those providers' terms before using this dataset to train a competing general-purpose model.