September 18, 2026 · 4 min read
Quartz Micro Preview Can Follow Instructions Now (Sort Of)
The full-parameter instruction-tuning pass on Quartz Micro Preview is done. Both checkpoints are published now — the base pretrain and this instruction-tuned release — and the eval result is real, but it's not a clean win. Here's the honest version.
What changed
Quartz Micro Preview started with zero instruction-following behavior — it was a raw pretrain, good at completing text and nothing else. Unlike Prism, Amethyst, and Copal (LoRA fine-tunes of an already-instruct-tuned base), teaching it to follow instructions at all needed a full-parameter supervised fine-tune, not a light adapter touch. That ran on databricks-dolly-15k, a human-written instruction dataset — 14,051 examples after filtering, 1,317 optimizer steps — on the same single GTX 1660 Ti that did the pretrain.
This was actually our second attempt. The first SFT pass underperformed, so this run used a higher learning rate and a fresh optimizer state rather than continuing from the first run's optimizer — a deliberate retry, not a resume.
The honest result
A modest, mixed improvement over the base checkpoint: clearly better at grounded, context-based QA (correctly extracts and restates facts from a given passage where the base model ignored the context entirely), but strict format- and length-following instructions still aren't reliably followed, and both checkpoints remain largely incoherent past the first sentence or two -- expected at this size, not a regression from the SFT pass.
Ten hand-authored prompts, not sampled from the training data, spanning open QA, closed/context QA, summarization, brainstorming, classification, extraction, general QA, instruction-following, and creative writing — the same held-out-eval discipline behind every release on this site.
Win — closed QA, uses given context correctly
Prompt: “Context: The Golden Gate Bridge is 2,737 meters long and was completed in 1937. Question: How long is the Golden Gate Bridge?”
Base ignores the context entirely and wanders into an unrelated story. The tuned model: “The Golden Gate Bridge is 2,737 meters long...” — correctly extracts and restates the fact.
No win — strict instruction-following
Prompt: “Write a short two-sentence greeting for a birthday card.”
The tuned model drifts into an unrelated riff about “the national party” and election seats — no better than the base model here, and no closer to a birthday greeting.
Both checkpoints also remain largely incoherent past the first sentence or two of most outputs. That tracks with the model's size (~1B total parameters, 394M active) — it's not a regression the SFT pass introduced, and it's not something a single instruction-tuning run was ever going to fix.
Why we're publishing this anyway
A model card that only shows the wins isn't an eval, it's marketing. The whole point of running a held-out comparison — on every release, not just the ones that flatter us — is that the result gets reported whichever way it lands. This one lands somewhere in the middle: real, narrow progress on grounded QA, and a clear reminder that a single SFT pass on a tiny model doesn't turn it into a general-purpose assistant.
Both checkpoints are on Hugging Face: VertexAGI/quartz-micro-preview-base and VertexAGI/quartz-micro-preview, each with the full architecture, training details, and the complete held-out transcript — not just the excerpts above. See the Research page for the full writeup.
— The Vertex AGI team