Why this model exists
Its Qwen3.5-0.8B twin, heed-tasks-only-qwen3.5-0.8b, scored slightly higher, but Google's official LiteRT-LM converter could not convert its hybrid attention (decision D103). Qwen3-0.6B can be converted.
EXP-038 tested whether the 0.6B gives up little enough quality to be worth the switch. It was trained on exactly the same data with the same settings:
- word-overlap F1 on the yardstick: 91.2 against 91.6;
- on real notes: 77.9 against 81.5.
Note: <the note text>Tasks:
The answer is {"tasks":["…","…"]}, then EOS. Use greedy decoding and no chat template.
Files
model.safetensors: bf16 weights for transformers.
heed-tasks-only-qwen3-0.6b-int8.litertlm: the int8 file that ran in the app (614,458,976 bytes, sha256 d17fa2d8d2779f8c88cd0b838a0dfa78e37a12af6d5bdd896d5b96957d2b3ab3).
- Made with
litert-torch export_hf … --cache_length=1024, which uses int8 weights and fp32 activations.
- Checked on a Mac on 30 notes before going on the phone.
- Batched and one-at-a-time answers were identical on 45 of 48 notes.
Training
Table with columns: Setting, Value| Setting | Value |
|---|
| Base | Qwen/Qwen3-0.6B-Base, full fine-tune, 596,049,920 trainable parameters |
| Data | The same 4,221 notes as the 0.8B twin, one row each. HeedSet (synthetic, Claude Opus), simulated user-style notes, notes I planned generated with ChatGPT, Claude brain dumps, 140 real notes from me and five friends and 10 typed by me |
| Labels | Older mixed labels: Sonnet-relabelled plus Sonnet-labelled HeedSet. They keep hedges like "Maybe…" |
| Recipe | lr 1e-5, AdamW, 5% warm-up + cosine, batch 8 × 2 = 16, 3 epochs (792 steps), max length 1,024, loss on answer tokens only, bf16 autocast, seed 0 |
| Hardware / time / cost | 1 × RTX 4090, 7.1 minutes, peak 13.9 GB; about $0.45 of GPU (estimate) |

Results (word-overlap F1 only; no judged score)
Validation has 371 notes; the yardstick is 199 of them (59 real plus 140 HeedSet).
- Yardstick F1 91.2 (recall 91.2, precision 91.2), with exact lists on 72.4% of notes.
- Real notes F1 77.9, with exact lists on 47.5%.
- HeedSet slice 95.2.
- Valid JSON on 199 of 199 notes.
On the phone (EXP-041: OnePlus CPH2717, Dimensity 8300, 40 test notes):
Table with columns: Metric, Value| Metric | Value |
|---|
| Time per note | 2.2 to 3.7 s |
| Prefill / decode | about 240 to 300 / 19 to 25 tokens per second (about 14 tokens per second with the screen locked) |
| Memory | PSS about 1.1 GB (616 MB of memory-mapped weights) |
| Heat | 36 to 40 °C over several minutes, using about 3.5 CPU cores |
| Background | It kept working with the screen locked and in the background |
What it taught the study
- The labels taught it to hedge. On the phone it wrote tasks like "Maybe call the bank", because about 1 in 5 hedged tasks in its training labels kept the hedge. That led to one consistent relabelling of every note with a single Sonnet prompt (EXP-046).
- It had no long notes. Adding 2,697 long synthetic notes later raised long-note quality by +9.4 (EXP-048/049).
- It had no buckets. The next models write a bucket for every task.
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.