mohitkrconnect/heed-voice-notes-to-tasks

Heed: voice notes to tasks

I built Heed, a voice-first note app that turns a spoken note into a short list of tasks, entirely on the phone. This dataset is the synthetic part of the data behind its model. Each row is one note, written the way people talk when they dump their head into a microphone, together with the task list a labeller made from it.

• Model trained on this data: https://huggingface.co/mohitkrconnect/heed-tasks-qwen3-0.6b
• Write-up of the whole study: https://huggingface.co/spaces/mohitkrconnect/heed-study

Every note in this dataset was written by a language model. There are no real notes here (see "What is not included").

What is in it

Split	Rows
train	7,669
validation	299
test	398
total	8,366

Rows by source (the source field):

source	What it is	train	validation	test
heedset	Notes about made-up people, written by Claude Opus	1,661	140	240
long_braindump	Long, scattered brain dumps (120 to 200 words), many from "ADHD-style" personas, written by Claude Sonnet 5.5	2,692	80	79
targeted	Short or tricky notes written on purpose to teach one skill each (self-correction, keeping numbers and names, rambling and returning to a topic, remarks with few or no tasks, one or two tasks in messy speech), written by Claude Sonnet 5.5	1,695	0	0
claude_braindump	Messy brain dumps in the voice of people from 16 English-speaking countries, written by Claude Opus and Claude Sonnet	479	79	79
simulated	Notes imitating one person's style, written by Claude Sonnet, Gemini and GPT-5.4 models	1,107	0	0
curated	35 natural-sounding notes I picked by hand from a larger set made by a template program	35	0	0

314 rows (3.8%) have an empty task list on purpose: the note holds only remarks or feelings and nothing to do. The median note has 83 words (range 2 to 209) and the mean list has 4.7 tasks. 929 rows use their own category list instead of the default one (3 to 10 categories).

How the notes were written

Language models wrote every note, each given a persona, a situation and a speaking style (clean, rambling, self-correcting, with filler words, with local words and place names). Models used: Claude Opus, Claude Sonnet 5.5, Gemini and GPT-5.4 models (the last two only for the simulated source). Notes that failed scripted format checks were dropped before labelling. Nothing was spoken: these are text notes written to look like what a speech-to-text system gives.

How the labels were made

All labels (tasks) in all three splits were made the same way, in two steps, both by Claude Sonnet 5.5 at temperature 0:

1. Labelling with one frozen prompt, labelling/labelling_prompt.txt. The labeller got the note and its category list. That prompt also asks for intent, a separate time field, a reminder flag and voice commands; this dataset keeps only the task text and the category, which is what the app model learns.
2. Putting the time back into the task. The spoken time words ("before Saturday", "at 5") were merged into the task text with labelling/time_rewrite_prompt.txt, 20 tasks per call. A rewrite was rejected if it lost any word of the task or the time, and the few that failed were read by hand.

The targeted notes were labelled with the same prompt plus a few rules for the error types they were written to cover (self-corrections, details to keep, remarks that are not tasks); the labeller never saw the plan the note was written from. Older labels from earlier stages of the project were all replaced by this one relabelling, so the dataset follows a single style. There was no per-note human check; I read samples.

Schema

One JSON object per line.

Field	Meaning
id	Opaque id such as heed-train-000123
note	The note text as written (punctuation and capitals kept)
note_plain	The same note lowercased, punctuation removed, apostrophes kept: re.sub(r"\s+", " ", re.sub(r"[^\w\s']", " ", note.lower())).strip(). This is what a speech-to-text system such as the one in the app produces
categories	The category list shown to the model: id, name, meaning
tasks	The label: a list of {"t": task text, "c": category id}
source	heedset, long_braindump, targeted, claude_braindump, curated or simulated
split	train, validation or test

Example row (synthetic):

{"id": "heed-train-000332",
 "note": "On Wednesday at haat buy 5 kg rice and mahua oil. Monday give beat register to ranger sir at range office.",
 "note_plain": "on wednesday at haat buy 5 kg rice and mahua oil monday give beat register to ranger sir at range office",
 "categories": [{"id": "tasks", "name": "General", "meaning": "general things to do"},
                {"id": "buy", "name": "Buy", "meaning": "things to purchase"},
                {"id": "watch", "name": "Watch", "meaning": "movies, shows, videos to see"},
                {"id": "ideas", "name": "Ideas", "meaning": "thoughts to remember"},
                {"id": "work", "name": "Work", "meaning": "office and job tasks"},
                {"id": "journal", "name": "Journal", "meaning": "feelings, emotions, stories with no action"}],
 "tasks": [{"t": "Buy 5 kg rice at haat on Wednesday", "c": "buy"},
           {"t": "Buy mahua oil at haat on Wednesday", "c": "buy"},
           {"t": "Give beat register to ranger sir at range office on Monday", "c": "work"}],
 "source": "heedset", "split": "train"}

To train like I did, build the prompt as Categories:\n- id (Name): meaning\n...\nNote: <note or note_plain>\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.