Same as the main model. The input is a raw prompt with the bucket list, then Note: <text>, then Tasks: and a newline. The output is {"tasks":[{"t":"…","c":"…"}]}. Use greedy decoding and no chat template.
Training
Table with columns: Setting, Value| Setting | Value |
|---|
| Base | Qwen/Qwen3-0.6B-Base, full fine-tune |
| Data | 8,610 notes × 2 forms (as written + lowercase/no punctuation) = 17,220 rows; labels by Claude Sonnet 5.5 (see main card, section 3) |
| Recipe | lr 1e-5, AdamW, 5% warm-up + cosine, batch 8 × 2, 3 epochs (3,231 steps), max length 1,024, loss on answer tokens only, bf16 autocast |
| Seed | 0 |
| Hardware / time | 1 × RTX 4090 24 GB, 32.96 minutes |
| Best validation loss | 0.341 after epoch 1. Epoch 3 is published because it scores better on tasks (see the main card, section 4.1) |

Results (validation only; this seed was not run on the sealed test)
All numbers are on the 451 validation notes, with the Gemini 3.1 Pro blind judge and the input in plain (speech-like) form.
Table with columns: Model, Judged F1, Precision, Recall, Real notes (59), Wrong items, Notes with no mistake| Model | Judged F1 | Precision | Recall | Real notes (59) | Wrong items | Notes with no mistake |
|---|
| This model (seed 0) | 77.5 | 78.8 | 76.2 | 81.9 | 81 | 50.6% |
| Seed 1 (the app model) | 79.4 | 81.2 | 77.6 | 86.2 | 76 | 51.9% |
| Previous recipe (EXP-054 seed 0) | 76.2 | — | — | 78.4 | 105 | 43.7% |
| Claude Sonnet 5.5 (label source) | 94.3 | — | — | — | — | — |

Word-overlap F1 against the Sonnet key (validation, epoch 3):
- Plain form: joint (task and bucket both right) 76.5, tasks 84.3, bucket 90.8. 49.4% of notes fully right; 100% readable JSON.
- Written form: joint 79.2.
What the targeted data did: average of seeds 0 and 1 minus the previous recipe, with 95% paired-bootstrap intervals.
- Real notes: +5.5 [+1.4, +10.3].
- Wrong items: from 99 to 78.5.
- Overall: +1.4 [−0.2, +2.9]. This failed the pre-registered primary rule and was recorded as a fail.
Seed-specific: on matched tasks, ideas recall fell from 94 to 82 for this seed, but stayed at 90 for seed 1.
Not measured for this model
- No sealed-test score.
- No phone conversion or phone timings. It has the same architecture and size as the app model, so it should behave the same, but that was not measured.
Limitations: the same as the main card.
Data: the synthetic part of the training, validation and test notes, with all five systems' answers and the judge's verdicts on the public test notes, is in mohitkrconnect/heed-voice-notes-to-tasks. My real notes and my friends' are not published.
License
Apache-2.0, the same as the base model. The training labels and most training notes were produced with Claude, Gemini and GPT models. Check those providers' terms before using this model to build a competing general-purpose model.