rockypod/ neotoi-coder:15b

57 1 month ago

A Rust / Dioxus 0.7 specialist fine-tuned via RAFT on Qwen3 (3 sizes), optimized for Tailwind v4 + WCAG 2.2 AAA accessibility and dioxus-primitives discipline. Built with CRAFT on a homelab RTX 3090 Ti. No cloud GPUs.

4b 8b 15b 
ollama run rockypod/neotoi-coder:15b

Details

1 month ago

61c317f4a078 · 9.0GB ·

model
qwen3
·
14.8B
·
Q4_K_M
template
{{- if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}<|im_start|>user {{ .Prompt }}<
system
You are Neotoi, an expert Rust and Dioxus 0.7 developer.
params
{ "num_ctx": 8192, "stop": [ "<|im_end|>" ], "temperature": 0.2 }

Readme

Neotoi Coder

A Rust / Dioxus 0.7 specialist LLM fine-tuned on 5,287 curated examples covering the full Dioxus 0.7 series (0.7.0–0.7.9), Tailwind v4, and WCAG 2.2 AAA accessibility. All three v3.2 variants are available.

Tags

  • rockypod/neotoi-coder:latest — 15B v3.2
  • rockypod/neotoi-coder:15b — 15B v3.2 (same as :latest)
  • rockypod/neotoi-coder:8b — 8B v3.2 (highest score)
  • rockypod/neotoi-coder:4b — 4B v3.2 (smallest / fastest)
  • rockypod/neotoi-coder:v3.1 — 15B v3.1 archive

v3.2 scorecards (114Q weighted exam, max 164.0)

All three variants clear the 90% publication bar and the 95% release bar.

Variant Score Weighted Raw
8B 97.56% 160.0 / 164.0 111 / 114
4B 97.56% 160.0 / 164.0 112 / 114
15B 95.12% 156.0 / 164.0 109 / 114

The 8B and 4B tie for the highest score. The 15B has the broadest Dioxus 0.7.4–0.7.9 surface coverage in its training.

15B

Tier Rate
T1 Fundamentals 100.0%
T2 RSX Syntax 100.0%
T3 Signal Hygiene 100.0%
T4 WCAG / ARIA 100.0% (was 78.6% in v3.1)
T5 use_resource 100.0%
T6 Hard Reasoning 100.0%
T7 Primitives + CSS 92.3%
T8 GlobalSignal / i18n 87.5%
T9 Static Navigator 100.0%
T10 Dioxus 0.7.4 100.0%
T11 Server Functions 100.0%
T12 Format Compliance 83.3%
T13 SyncStore 0.0% (2-question tier, single miss)

8B

Tier Rate
T1–T7 100.0% each
T8 GlobalSignal / i18n 87.5%
T9–T12 100.0% each
T13 SyncStore 50.0% (2-question tier, single miss)

4B

Tier Rate
T1–T11 100.0% each
T12 Format Compliance 66.7%
T13 SyncStore 100.0%

Pick by hardware

  • 15B (8.4 GB) — full Dioxus 0.7.4–0.7.9 surface
  • 8B (4.7 GB) — highest score, ~40% faster than 15B
  • 4B (2.3 GB) — disk / RAM constrained, perfect SyncStore

What’s new in v3.2

  • Full Dioxus 0.7 series (0.7.0–0.7.9): Scoped CSS + CSS modules (0.7.3), SyncStore + use_store_sync (0.7.2), onauxclick / onscrollend events (0.7.3), server-only extractors with serde_qs, inert attribute, web panic resilience, IntoAttributeValue for &T, Action::PartialEq, iOS widget bundling, WebSocket Stream + Sink, FFI for Kotlin/Java/Swift.
  • T4 WCAG/ARIA: 78.6% → 100% on 15B — dedicated correction training ensures rsx! is never dropped on accessibility-heavy components.
  • Format compliance: fenced-code-only outputs, no orphan </think>, no prose preamble.
  • Preserve-and-append: edits to .ftl catalogs, Cargo.toml, and route enums add to existing files rather than replacing them.
  • 5,287 training examples across 57 topics (up from 4,880 / 43 in v3.1).

Base models

  • 15B: Qwen3-Coder-14B (14.8B params)
  • 8B: Qwen3-8B (8.2B params)
  • 4B: Qwen3-4B (4.0B params, tied embeddings)

License

Neotoi Coder Community License — see HuggingFace repo for full text.

Links