VertexAGI/prism-caption-1-5-micro

Prism Caption 1.5 Micro

A 0.6B-parameter model that does one thing: turn a user's first message into a short chat title.

Part of the Prism family of small, single-purpose models. Successor to Prism Caption 1 Micro (Gemma-3-1B, 339 examples).

Overview

Given the first message in a new conversation, this model returns a short, specific title — the "New chat" auto-naming behaviour that ChatGPT, Claude and similar assistants provide. Nothing else. At 0.6B parameters it is small and fast enough to run this narrow task locally instead of spending a frontier-model call on it.

Results

Evaluated on 24 held-out topics that appear nowhere in the training data (the v1 evaluation sampled from its own training topic list, so it measured memorisation rather than generalisation — that mistake is corrected here).

Chat titling has no single correct answer, so the metrics are behavioural:

Metric	Base Qwen3-0.6B	Prism Caption 1.5
Format issues (preamble, punctuation, multiline)	10/24	0/24
Within the 3-6 word spec	14/24	23/24
Shares content words with the message	24/24	21/24
Average length	5.8 words	3.3 words

The win is format reliability. The base model leaks the prompt back ("What's the deal with picking a beginner-friendly board game for si..."), keeps question marks ("Sourdough starter smells like acetone?"), and echoes the user's phrasing instead of titling it. The tuned model does not do these things.

The honest cost: it is terser, and sometimes too terse. Relevance fell from 24/24 to 21/24 because a few titles over-compress and drop the actual subject:

Message	Title
choosing a mattress for side sleepers with shoulder pain	"Shoulder Sleep Matters"	loses mattress
why my sourdough starter smells like acetone	"Acetone Smell Troubleshooting"	loses sourdough

Most titles are good — "Last Minute Flight Cancellation", "First Half Marathon Training", "Laptop Lid Residue Removal", "Car Braking Grinding Noise" — but if your UI needs the subject noun preserved, test on your own traffic first.

Training

Base	Qwen/Qwen3-0.6B (0.6B parameters, via mlx-community/Qwen3-0.6B-4bit)
Method	LoRA, rank 8, scale 20, 8 layers
Optimizer	Adam, lr 1e-5 constant, batch 4, seq len 256
Steps	1,500 iters; released checkpoint is iter 1,500 (val loss 0.209)
Framework	MLX (mlx-lm) on Apple Silicon
Data	1,000 examples over 167 topics and 20 phrasing templates (900 train / 100 valid)

Teacher. Titles were distilled through NVIDIA NIM. Generation began on google/diffusiongemma-26b-a4b-it, which produced good titles but whose NIM rate limit is far tighter than the nemotron endpoints — 82% of requests returned 429. The remaining ~957 examples were generated with nvidia/nemotron-3.5-lightning-30b-a3b. The corpus is therefore mixed-teacher: roughly 43 examples from diffusiongemma, the rest from Lightning.

On topic breadth. v1 used 51 topics for 339 examples. Scaling that to 1,000 would mean ~20 repeats per topic; narrow template pools cause the model to memorise the pool instead of learning the task. The bank was widened to 167 topics first, so each recurs ~6 times. The result set contains 714 unique titles across 1,000 examples.

Limitations

• This is a 0.6B-parameter model — small enough to run cheaply and locally, but limited in general capability outside its one narrow task.
• Over-compression. Averages 3.3 words against a 3-6 word spec, and occasionally drops the subject noun (see above).
• English only.
• First messages only. It was not trained to re-title a conversation from later turns or from full transcripts.
• Not a summariser. It produces labels, not descriptions.
• Inherits Qwen3-0.6B's knowledge cutoff and general limitations.

Usage

MLX

from mlx_lm import load, generate
import re

model, tok = load("VertexAGI/prism-caption-1-5-micro")

SYSTEM = ("You name chat conversations. Given the user's first message, reply with ONLY a "
          "short, specific chat title (3-6 words, title case, no quotes, no punctuation at "
          "the end, no preamble). Nothing else -- just the title.")

msgs = [{"role": "system", "content": SYSTEM},
        {"role": "user", "content": "Hey, why does my car make a grinding noise when braking?"}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False,
                               enable_thinking=False)
out = generate(model, tok, prompt=text, max_tokens=24, verbose=False)
print(re.sub(r"<think>.*?</think>", "", out, flags=re.S).strip())
# -> Car Braking Grinding Noise

Qwen3 emits an empty <think></think> block even in non-thinking mode — strip it before using the output.

GGUF (llama.cpp)

Use llama.cpp's own chat template rather than hand-building the prompt; a hand-rolled template makes the model emit a thinking preamble and degrades the title.

llama-cli -m prism_caption_1_5_Q4_K_M.gguf \
  -sys "You name chat conversations. Given the user's first message, reply with ONLY a short, specific chat title (3-6 words, title case, no quotes, no punctuation at the end, no preamble). Nothing else -- just the title." \
  -p "Hey, why does my car make a grinding noise when braking?" \
  -n 20 --temp 0 -st --chat-template-kwargs '{"enable_thinking":false}'
# -> Car Braking Noise

Formats

Format	File	Size	Notes
MLX (4-bit)	model.safetensors + config	335 MB	Apple Silicon via mlx-lm
GGUF (Q4_K_M)	prism_caption_1_5_Q4_K_M.gguf	378 MB	llama.cpp, LM Studio, Ollama

License

Apache 2.0, inherited from Qwen3.