Why this model exists
EXP-050 asked which output format is better when one 0.6B model writes task, bucket, time and reminder. A line format is about 30% shorter than JSON, which matters on a phone CPU where decoding dominates time. The test was run with everything else held fixed: same data, same seed, same four epochs, same recipe. Only the answer format differs. The paired JSON model is heed-full-output-json-4ep-qwen3-0.6b.
The design set a margin before the run: lines had to be within 1 point of JSON to be kept as a serious option. Prediction P50.2 said it would be within 2 points. It missed by 0.04.
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 one task per line, then EOS. An invented example:
Call the bank|tasks|at 5|rBuy milk|buy
Fields are text|bucket|time words|r. Trailing empty fields are dropped, so a task with no time and no reminder is text|bucket. The reminder field is the letter r when a reminder is needed. A note with no tasks is a single -.
Mean output tokens, older notes and long notes: lines 45 and 89, short JSON 65 and 126, full-word JSON 76 and 159, tasks only 34 and 65.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "mohitkrconnect/heed-full-output-lines-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 evaluatedtext = tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True) tasks = []for line in text.splitlines(): if line.strip() in ("", "-"): continue parts = line.split("|") tasks.append({"t": parts[0], "c": parts[1] if len(parts) > 1 else None, "w": parts[2] if len(parts) > 2 and parts[2] != "r" else "", "r": parts[-1] == "r" and len(parts) > 2})print(tasks)
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_lines_s0, epoch 4 of 4 (epoch 3 scored 69.0) |
| 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, 16.56 minutes, best validation loss 0.4131 (step 866), peak 16.4 GB |
| Hardware / cost | 1 × RTX 4090. About 0.25ofGPUforthisrunatthequoted0.90 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, Lines (this model), JSON, same recipe| 451 validation notes | Lines (this model) | JSON, same recipe |
|---|
| Joint F1 | 69.4 | 71.5 |
| Task F1 | 84.7 | 86.2 |
| Bucket, on matched tasks | 91.9 | 92.3 |
| Reminder F1 | 90.7 | 91.2 |
| Time F1 | 93.0 | 93.8 |
| Notes fully right | 44.1% | 45.9% |
| Real notes (59), joint F1 | 59.6 | 65.2 |
| Long notes (80), joint F1 | 67.9 | 68.7 |
| Readable answers | 99.8% (real notes 98.3%) | 100% |
| Best validation loss | 0.4131 | 0.2704 |
The gap of 2.04 points on joint F1 is over the 1-point margin fixed in the design, and the lines format was also slightly less often readable. On the 59 real notes the gap is larger (59.6 against 65.2), but with seed noise of this size real-note differences are not claims. Task F1 is 1.5 points lower, so the loss is mostly in finding and wording the task, not in the extra fields.
Phone speed for this model: not measured. The format is about 30% fewer tokens, so it stays a latency candidate; no test showed the saving is worth the quality loss.
What it fed
D120: JSON beats lines by 2 points; JSON is the format from then on, and it is the format of the app model today. The full set of EXP-050 variants (JSON warm start, 8-epoch runs) is in the chosen JSON model's card.
Limitations
- One seed per format, so the 2-point gap rests on a single pair of runs. The study's own seed noise is about 3 points on other metrics (D101), so the gap is suggestive, not proven.
- Quick word-overlap scores against one model's labels; never judged, never on the sealed test.
- No lowercase or unpunctuated training data.
- 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.