Instructions to use mohitkrconnect/heed-tasks-qwen3-0.6b-exp054 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use mohitkrconnect/heed-tasks-qwen3-0.6b-exp054 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="mohitkrconnect/heed-tasks-qwen3-0.6b-exp054")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("mohitkrconnect/heed-tasks-qwen3-0.6b-exp054")
model = AutoModelForCausalLM.from_pretrained("mohitkrconnect/heed-tasks-qwen3-0.6b-exp054", device_map="auto")

LiteRT-LM

How to use mohitkrconnect/heed-tasks-qwen3-0.6b-exp054 with LiteRT-LM:

# LiteRT-LM runs on various platforms (Android, iOS, Windows, Linux, macOS, IoT, Web/WASM)
# and supports many APIs (C++, Python, Kotlin, Swift, JavaScript, Flutter).
# For platform-specific integration guides, please refer to the official developer website:
# https://ai.google.dev/edge/litert-lm

# To try LiteRT-LM, the easiest way is to use our CLI tool.
# 1. Install the LiteRT-LM CLI tool:
pip install -U litert-lm

# 2. Download and run this model locally:
# See: https://ai.google.dev/edge/litert-lm/cli
# A single .litertlm file in the repo is picked automatically; otherwise the CLI asks which one to run
# (or pass its name right after the repo id).
litert-lm run \
  --from-huggingface-repo=mohitkrconnect/heed-tasks-qwen3-0.6b-exp054 \
  --prompt="Write me a poem"

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use mohitkrconnect/heed-tasks-qwen3-0.6b-exp054 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "mohitkrconnect/heed-tasks-qwen3-0.6b-exp054"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mohitkrconnect/heed-tasks-qwen3-0.6b-exp054",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/mohitkrconnect/heed-tasks-qwen3-0.6b-exp054

SGLang

How to use mohitkrconnect/heed-tasks-qwen3-0.6b-exp054 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "mohitkrconnect/heed-tasks-qwen3-0.6b-exp054" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mohitkrconnect/heed-tasks-qwen3-0.6b-exp054",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "mohitkrconnect/heed-tasks-qwen3-0.6b-exp054" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "mohitkrconnect/heed-tasks-qwen3-0.6b-exp054",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use mohitkrconnect/heed-tasks-qwen3-0.6b-exp054 with Docker Model Runner:
```
docker model run hf.co/mohitkrconnect/heed-tasks-qwen3-0.6b-exp054
```

Heed tasks, Qwen3-0.6B, EXP-054 (the version before targeted data)

This model was the Heed app model before the current one. It was then replaced by heed-tasks-qwen3-0.6b, which adds 1,696 targeted training notes. It is the baseline for that change, published with its phone file.

For the full story (problem, judge, sealed-test results, limitations), read the main model card.

What EXP-054 changed, and why

The previous model (EXP-050) also wrote a reminder flag and a separate time field. A phone test then showed two problems:

On the lowercase, unpunctuated text that comes out of the speech model, its quality fell by 5.7 points (−5.7 [−11.4, −0.7]).
The extra fields cost quality.

EXP-054 made two changes:

The answer became tasks plus a bucket only: {"tasks":[{"t":"…","c":"…"}]}. Any spoken time is written inside the task text, and reminders are set by hand in the app (decision D134).
Every note was trained twice: as written, and lowercase with no punctuation.

Result on the phone: the punctuation drop fell from 5.7 to 0.2 points (−0.2 [−4.6, +4.7]).

Files

model.safetensors: bf16 weights for transformers.
heed-tasks-qwen3-0.6b-exp054-int8.litertlm: the int8 LiteRT-LM file that ran in the app (614,458,976 bytes).

The prompt format is identical to the main model; see How to use it.

Training

Setting	Value
Base	`Qwen/Qwen3-0.6B-Base`, full fine-tune
Data	6,914 notes × 2 forms = 13,828 rows: 4,217 older notes (synthetic HeedSet, notes I planned, simulated, Claude brain dumps, 150 real notes from me and five friends) plus 2,697 long synthetic brain dumps
Labels	All by Claude Sonnet 5.5, under one frozen prompt. Sonnet rewrote 10,752 tasks to keep the spoken time inside the text; a rewrite was rejected if it lost a word, and the 53 that failed that check were read by hand
Recipe	lr 1e-5, AdamW, 5% warm-up + cosine, batch 8 × 2, 3 epochs (2,595 steps), max length 1,024 (longest row 543 tokens), loss on answer tokens only, bf16 autocast, seed 0
Hardware / time / cost	1 × RTX 4090 24 GB, 25.8 minutes; about $1.40 of GPU for the whole experiment (estimate)

Results

Validation: 451 notes, Gemini 3.1 Pro blind judge, plain input.

Judged F1 76.2; real notes (59) 78.4.
By length: under 15 words 87.1, 15–39 words 78.2, 40–99 words 72.6, 100+ words 77.0.
Against the earlier EXP-050 model on the same notes: +4.8 [+2.4, +7.4].
Claude Sonnet 5.5, which wrote the labels, scores 94.3 on these notes.
A second judge, GPT-5.4, on 200 notes: 72.5 against Sonnet's 88.9.

On the phone: OnePlus CPH2717 (Dimensity 8300), int8, 109 notes.

Word-overlap joint F1: 75.3 on plain input vs 75.5 on written input.
Bucket accuracy 94.3.
82.1% of spoken times kept.
Median 3.2 s per answer, p95 6.8 s.
All answers were readable JSON, with no repeated tasks.

Error review (EXP-055, 561 mistakes on 451 validation notes):

Kind of mistake	Share
Task dropped	22%
Wrong bucket	19%
Invented task	18%
Reworded badly	15%
Time lost	13%
Wrongly split	8%
Wrongly merged	5%

Notes with no mistake fall from 80% under 15 words to 26% at 40–99 words. This review is what the targeted data of the next version was built to fix.

Not measured: there is no sealed-test score for this model. Every number above comes from validation or the phone check.

Limitations: the same as the main card, and it is weaker on long, rambling notes.

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.