simplixity-title-0.8b

Qwen3.5-0.8B fine-tuned to name chats in Simplixity, a self-hosted research assistant. Given the first message of a conversation in English or German, it writes an English title of at most six words.

GGUF files for LM Studio and llama.cpp are in simplixity-title-0.8b-GGUF.

Use with Ollama

From a download of this repo:

ollama create simplixity-title:0.8b-v1-q8_0 -f Modelfile --quantize q8_0

In Simplixity, turn on chat titles in the settings and choose simplixity-title:0.8b-v1-q8_0 with provider Ollama. --quantize q4_K_M gives a 728 MB model for small machines, at a small cost in length discipline (see below).

Prompt

The model was trained on one prompt and should get exactly that one. System message:

Write a title of at most six words for a conversation that begins with this
message. Describe the subject. No quotation marks, no trailing punctuation, no
"chat about". Reply with the title only.

Write this in English, whatever language the question and the sources are in.
This is interface text (what the app shows about its own work) and not the
answer, which is written separately and follows the reader's language.

User message: the first message of the chat, cut to 1,000 characters. Simplixity asks for at most 24 tokens at temperature 0.2. Thinking does not need to be turned off, because the model closes Qwen's empty thinking block itself.

Results

60 first messages, 30 German and 30 English, three runs each, sent the way Simplixity sends them.

	Qwen3.5-0.8B	This model, Ollama Q8_0	This model, Ollama Q4_K_M
All format rules (six words, English, no quotes, no trailing punctuation)	35%	95%	87%
Key term of the message named	82%	89%	91%
Usable, sorted by hand	35%	75%	not sorted
Answers the question instead of naming it	22%	0%	not sorted
Median per title (M3 Pro)	0.14 s	0.32 s	0.30 s

Every format miss is a title with seven words. The full write-up is in the Simplixity repo.

Limits

• German words the 0.8B does not know come out wrong: "Rente" becomes "Rental", "Süßstoff" becomes "Sugar", "Siebträgermaschine" becomes "Sieve Jar Machine".
• About one German message in five gets a German title.
• Training stopped at 0.84 epochs (checkpoint 1,500 of 3,562) after repeated GPU hangs on the Mac. A longer run may do better.

Training

• Data: 2,057 synthetic first messages and their titles, written by Claude Haiku 5.5 through Simplixity's title prompt. Titles that broke a rule went back to Haiku once with the rule named. 1,874 passed every check: training/train.jsonl (1,781) and training/valid.jsonl (93), in the {"messages": [...]} chat format.
• Method: LoRA, rank 16 on every layer, loss on the title only, batch 1 with gradient accumulation 8, learning rate 2e-5, mlx-lm 0.32. Config: training/lora-v1.yaml.
• Merge: mlx_lm fuse saves MLX's own layout (renamed tensors, transposed conv1d, 1.0 added to every RMSNorm weight). training/export_hf.py copies the tuned linear weights back into the original checkpoint instead, so these weights are in the standard Qwen3.5 layout, vision tower included and unchanged.

Licence

Apache 2.0, like the base model.