Why this model exists
EXP-050 compared two answer formats for a model that writes task, bucket, time and reminder: JSON and one pipe-separated line per task. A fair test needs the two runs to differ only in format. This is the JSON run that matches the lines model: same data, same seed (0), same 4 epochs, same recipe.
Result: JSON 71.5 joint F1 against 69.4 for lines, a gap of 2.04 over the 1-point margin fixed in the design (decision D120).
The model picked for the phone test was a different run, the 8-epoch seed-1 JSON model. Its joint F1 is 71.4, the same within noise, so this 4-epoch model is as good in practice and about half the training time.
Raw text, no chat template, no system prompt. The prompt ends with a newline after Tasks:.
Categories:- tasks (General): general things to do- buy (Buy): things to purchase- watch (Watch): movies, shows, videos to see- ideas (Ideas): thoughts to remember- work (Work): office and job tasks- journal (Journal): feelings, emotions, stories with no actionNote: <the note text>Tasks:
The answer is compact JSON followed by EOS. An invented example:
{"tasks":[{"t":"Call the bank","c":"tasks","w":"at 5","r":1},{"t":"Buy milk","c":"buy"}]}
t is the task text, c the bucket id, w the spoken time words and r is 1 for a reminder. Empty w and r are left out. A note with nothing to do gives {"tasks":[]}.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "mohitkrconnect/heed-full-output-json-4ep-qwen3-0.6b"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16) CATEGORIES = """Categories:- tasks (General): general things to do- buy (Buy): things to purchase- watch (Watch): movies, shows, videos to see- ideas (Ideas): thoughts to remember- work (Work): office and job tasks- journal (Journal): feelings, emotions, stories with no action"""note = "call the bank at five and buy milk on the way home"prompt = f"{CATEGORIES}Note: {note}\nTasks:\n" # raw prompt, no chat template ids = tok(prompt, return_tensors="pt").input_idsout = model.generate(ids, max_new_tokens=512, do_sample=False) # greedy, as evaluatedprint(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
Weights are bf16. The duplicated tied lm_head was removed before upload (596,049,920 parameters).
Training
Table with columns: Setting, Value| Setting | Value |
|---|
| Run | EXP-050 f_json_s0, epoch 4 of 4 (epoch 3 scored 69.7) |
| Base | Qwen/Qwen3-0.6B-Base, full fine-tune, 596,049,920 parameters |
| Data | 6,914 notes, one row each, no punctuation variants. Validation: 451 notes. The 617 sealed-test notes were never put on the training machine |
| Labels | Claude Sonnet 5.5 under one frozen labelling prompt, temperature 0. Bucket, time words and reminder flag come from those labels |
| Recipe | lr 1e-5, AdamW (weight decay 0), 5% warm-up then cosine, gradient clipping 1.0, batch 8 × 2 = 16, max length 1,024, loss on answer tokens only, bf16 autocast, gradient checkpointing, seed 0 |
| Steps / time | 1,732 steps, 18.2 minutes, best validation loss 0.2704 (step 866), peak 17.6 GB |
| Hardware / cost | 1 × RTX 4090. About 0.27ofGPUforthisrunatthequoted0.90 per hour (derived estimate, not an invoice) |

Training notes (6,914): 1,108 simulated notes in my style (written by Claude Sonnet 5.5, Gemini and GPT-5.4); 1,662 HeedSet notes (synthetic, written by Claude Opus; 1,263 labelled earlier by Sonnet 5.5 and 399 relabelled in EXP-046); 781 notes planned by me and generated with ChatGPT; 480 Claude brain dumps (220 Opus, 260 Sonnet); 36 curated template notes; 140 real notes from me and five friends; 10 notes typed by me; 2,697 long synthetic notes written by Sonnet 5.5. English only.
Results
Read this first. These are quick word-overlap scores against Claude Sonnet 5.5's labels on the 451 validation notes, as written. A task counts only when its text matches by word overlap and its bucket, reminder and time are all right ("joint F1"). They are not comparable with the blind-judge numbers on the app model card. This checkpoint was never judged and never run on the sealed test.
Table with columns: 451 validation notes, This model (JSON, 4 epochs)| 451 validation notes | This model (JSON, 4 epochs) |
|---|
| Joint F1 | 71.5 |
| Task F1 | 86.2 |
| Bucket, on matched tasks | 92.3 |
| Reminder F1 | 91.2 |
| Time F1 | 93.8 |
| Notes fully right | 45.9% |
| Real notes (59), joint F1 | 65.2 |
| Long notes (80), joint F1 | 68.7 |
| Readable answers | 100% |
Pair with lines (same recipe, only the format differs): joint 69.4, task 84.7, bucket 91.9, reminder 90.7, time 93.0, fully right 44.1%, readable 99.8%.
A caution about the last epoch. Here the 4-epoch run peaked at its last epoch while validation loss had already turned up after epoch 2 (0.270 to 0.400). A cosine schedule flatters the last epoch, so 4 epochs did not prove convergence. Two 8-epoch runs then showed a plateau from about epoch 2 or 3, and eight epochs added nothing (D122). The 4-epoch seed-1 run was dropped for budget.
Phone behaviour for this checkpoint: not measured. The other EXP-050 checkpoints and the phone test are described on the chosen JSON model's card.
What it fed
- D120: JSON beats lines by 2 points.
- Together with the 8-epoch runs, it showed the model needs about 2 to 4 epochs (D122).
- Both formats were later set aside by D134, which dropped reminder and the separate time field.
Limitations
- Quick word-overlap scores against one model's labels; never judged, never on the sealed test.
- Reminder and time fields are weak on real notes, and real-note scores swing by several points between seeds.
- No lowercase or unpunctuated training: such input costs this family about 5.4 joint F1 on the phone (EXP-052).
- Trained on 150 real notes from me and five friends; memorisation was not tested.
- English only.
Links
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.