taleef/jobtrail-field-notes

JobTrail field notes: synthetic training set v1

799 tradesperson end-of-job notes, each paired with a structured job record, for fine-tuning a small on-device model to turn a note into a record. It's the training set of JobTrail (issue #74). The frozen eval sets it was kept apart from are in the same repository (data card).

All of it is synthetic. Gemini 3.8 Flash wrote and labeled every note, through the Antigravity CLI. No real customer or job is in it. Google's terms for Gemini output apply in addition to the MIT license, and OpenAI's to the 150 GPT-6 Luna cross-check labels in the raw configuration.

Configurations

• default (train.jsonl, 799 rows): the training set. Fields are id, note, target (the record), trade, style, tags and source.
• raw (raw.jsonl, 930 rows): every note of plans 1–930, kept or not, with its plan, the labels (labels, keyed by the models that labeled the note; null would mean an unparseable or invalid answer, and none is), kept, and why it was dropped (reason, contradicts). It's there to audit the filter or to build a different one.

target follows schema v2: jobType, workPerformed, issuesFound, materials (name, quantity, unit), laborMinutes, customerApproved and followUps. A value the note doesn't state is null or empty. The labeling rules are in data/LABELING.md.

Other files: filter.json (counts, mix and the SHA-256 of every data file), leakage.json, heldout.json, config.yaml (every setting of the run) and raw/: the files the run appended to, from which filter rebuilds everything above. They include every model answer and the time and tokens of every call, and 30 notes written past plan 930 that v1 doesn't use.

How it was made

1. Plan. A seeded sampler draws a record from the job templates in data/scenarios.yaml: 9 trades, 4 note styles (terse, rambling, spoken-fillers, run-on) and 8 hard-case tags (hours phrasing, negation, self-correction, multiple materials, no materials, supply-house trip, extra labor, approval absent). The tags that caused most baseline failures are weighted up, and no plan repeats an eval answer key. A plan may repeat in a different style: 7 plans appear twice among the kept notes. The plans are a fixed prefix (1–930) of a 6,000-plan pool, so a later, larger run adds notes without changing these.
2. Write. gemini-3.8-flash-medium writes each note from its plan in its own words, 15 at a time.
3. Label, blind. Two labelers read only the note: gemini-3.8-flash-high (97.3% zero-edit on the dev set with this prompt) and, separately, gemini-3.8-flash-medium. The target is the High label, so it uses the note's wording, not the plan's.
4. Filter. A note is kept only if:
   • both labels are valid;
   • they agree, under the JobTrail scorer's zero-edit test in both directions;
   • the target matches what the plan fixed: the trade, minutes and approval; the number of tasks, issues and follow-ups (not their wording); and the materials by quantity and unit (not by name). Up to one extra task per material is allowed, since the eval gold also lists narrated installs as tasks. One planned issue, "Doorknob punched through the drywall", may be left out, as the eval gold labels it both ways (LABELING.md reads it as the job's own task); 3 kept notes leave it out.
5. Leakage check. A kept note is dropped if it's a near-copy of any test or dev note: 5-word shingle Jaccard ≥ 0.2, or 8-gram containment ≥ 0.3. Both thresholds were set from the test-vs-dev distribution (95th percentiles 0.139 and 0.154) before any training note existed.

Results

Step	Notes
Planned and written	930
Labelers disagree	−77
Target contradicts the plan	−43
Near-copy of an eval note	−11
Kept	799 (85.9%)

Of the 43 contradictions, the fields involved were materials 21, issues 18, tasks 18, trade 15, minutes 15, approval 12 and follow-ups 11 (a note can break several).

Kept notes by style: terse 203, rambling 182, spoken-fillers 204, run-on 210. By trade: 74–109 each. Every hard-case tag has at least 86 kept notes; self-correction, the most common baseline failure, has 188. The full mix is in filter.json.

Cross-check, not part of the filter. GPT-6 Luna (another model family, through the Codex CLI) labeled the first 150 notes blind. On the 126 of them that were kept, its label agrees with the target on 83 (66%) and keeps the plan facts on 98 (78%).

Held-out templates

One job template per trade never reaches training. That way a fine-tuned model can be scored separately on eval notes from templates it has never seen, to tell whether it learned the task or the templates. The slice is 49 of the 163 test notes and 27 of the 110 dev notes; the ids are in heldout.json.

Trade	Held-out template
appliance	Replaced dishwasher drain pump
carpentry	Rehung sagging exterior door
cleaning	Deep cleaned kitchen
electrical	Installed GFCI outlets in bathroom
general	Assembled flat-pack wardrobe
hvac	Replaced blower motor capacitor
painting	Stained deck
plumbing	Replaced toilet wax ring
roofing	Cleaned gutters

Limitations

• Smaller than planned. The issue asked for 3,000–5,000 kept notes. At about 43 s of model time per note (writing 7 s, Flash High labeling 22 s, Flash Medium 13 s, from the usage logs), the owner stopped at plans 1–930 and will add more only if fine-tuning shows it needs them.
• One model family. The writer and both labelers are Gemini 3.8 Flash, so the targets carry its reading of the labeling rules. The cross-check shows another family reads about a third of the notes differently.
• Agreement isn't correctness. Two labelers can agree on the same mistake. The plan check catches wrong numbers, trades and counts, but not wrong wording, material names or field routing.
• Templates. Outside the held-out ones, training and eval notes share job templates, and 442 of the 810 notes checked share at least one 8-gram with an eval note (54 of the 110 dev notes do with test, too). Only near-copies were dropped.
• No human reviewed it.

Reproduce

From the repository, in ml/:

uv run python scripts/gen_train_notes.py write
uv run python scripts/gen_train_notes.py label gemini
uv run python scripts/gen_train_notes.py label gemini-medium
uv run python scripts/gen_train_notes.py filter

The limit (930) comes from config.yaml. Every stage resumes where it stopped, and --shard K/N splits a stage across processes. Model output isn't deterministic, so a rerun writes different notes. Put this dataset's raw/ files in data/train/raw/ and filter rebuilds train.jsonl and raw.jsonl byte for byte; filter.json records the SHA-256 of both and of every raw file.