---
license: other
license_name: neotoi-coder-community-license
language:
- en
- vi
base_model: Qwen/Qwen3-Coder-14B
tags:
- dioxus
- rust
- accessibility
- wcag
- fine-tuned
- raft
- code
- server-functions
pipeline_tag: text-generation
---

# Neotoi Coder v3.0

A Rust/Dioxus 0.7 specialist fine-tuned from Qwen3-Coder-14B using RAFT
(Retrieval-Augmented Fine-Tuning). Expanded for Dioxus 0.7.3–0.7.5:
scoped CSS, new event handlers, real WebSocket Stream+Sink, GlobalSignal
cache rebuilds, and fullstack server functions — on top of v2.0's
Tailwind v4 + WCAG 2.2 AAA + i18n coverage.

## What's New in v3.0

- **New Tier 11 — Server Functions:** Clean sweep 4.5/4.5.
  `#[server]` with server-only extractors, fullstack WebSocket
  one-line syntax, `ServerFnError` with custom HTTP status codes
- **Scoped CSS (0.7.3):** `css!()` macro for inline scoped styles,
  native `.module.css` imports
- **New event handlers (0.7.3):** `onauxclick` (middle-click),
  `onscrollend` (scroll-end detection)
- **WebSocket Stream+Sink (0.7.4):** Real `stream.next()` and
  `sink.send()` — no more `tokio::sleep` simulation (v2.0 gap closed)
- **GlobalSignal cache rebuild:** Idiomatic `.write()` on pre-rsx! lets
- **v2.0 regression fixes:** Q8 (button `r#type:`), Q16 and Q21
  (RSX details/summary fidelity)
- **Dataset:** 4,535 curated examples — 4,185 v2.0 base plus 350
  new cross-topic pairs covering the surface above

## Exam Results

### v3.0 — 103 Question Weighted Exam

| Tier | Questions | Weight | Score | Max | Status |
|---|---|---|---|---|---|
| T1 Fundamentals | Q1–12 | 1.0 | 11.0/12 | 12 | ✅ |
| T2 RSX Syntax | Q13–24 | 1.0 | 8.0/12 | 12 | ⚠️ Regression |
| T3 Signal Hygiene | Q25–36 | 1.0 | 9.5/12 | 12 | ✅ |
| T4 WCAG/ARIA | Q37–50 | 1.5 | 19.5/21 | 21 | ✅ |
| T5 use_resource | Q51–58 | 1.5 | 12.0/12 | 12 | ✅ Perfect |
| T6 Hard Reasoning | Q59–68 | 2.0 | 15.0/20 | 20 | ✅ |
| T7 Primitives+CSS | Q69–80 | 1.5 | 15.0/18 | 18 | ✅ |
| T8 GlobalSignal/i18n | Q81–88 | 1.5 | 10.5/12 | 12 | ✅ |
| T9 Static Navigator | Q89–94 | 1.5 | 9.0/9 | 9 | ✅ Perfect |
| T10 Dioxus 0.7.4 | Q95–100 | 2.0 | 10.0/12 | 12 | ✅ |
| T11 Server Functions | Q101–103 | 1.5 | 4.5/4.5 | 4.5 | ✅ Perfect |
| **Overall** | **Q1–103** | | **124.0/144.5** | **144.5** | **✅ 85.8%** |

Release threshold: 85% (123.0/144.5). v3.0 clears it with 1.0 point to spare.

### Version History

| Version | Score | Exam | Status |
|---|---|---|---|
| v1.0 | 51/60 (85.0%) | 60Q standard | Published |
| v2.0 | 135.5/140 (96.8%) | 100Q weighted | Published |
| v3.0 | 124.0/144.5 (85.8%) | 103Q weighted | Published |

v3.0's overall percentage is lower than v2.0 because the 103Q exam is
meaningfully harder: T11 Server Functions was added, T6 Hard Reasoning
and T10 Dioxus 0.7.4 got new items, and T2 RSX Syntax has regressed
against v2.0. The clean sweeps on T5 / T9 / T11 and the perfect fixes
at Q8 / Q16 / Q21 reflect the intended v3 surface expansion.

### Improvements Over v2.0

- **Q8** — button `r#type:` attribute precision — fixed
- **Q16** — RSX details/summary fidelity — fixed
- **Q21** — semantic tag preservation — fixed
- **T11 clean sweep** — 4.5/4.5 on the brand-new Server Functions tier
- **T5 use_resource** — held perfect (12.0/12)
- **T9 Static Navigator** — held perfect (9.0/9)

### Known Regressions — v3.1 Targets

T2 RSX Syntax dropped from 10/12 (v2.0) to 8.0/12 (v3.0):

- **Q14** — RSX attribute placement precision
- **Q15** — RSX attribute placement precision
- **Q22** — RSX attribute placement precision

Root cause under investigation. Targeted for v3.1.

## Model Details

- **Base model:** Qwen3-Coder-14B (fresh base — never fine-tune a fine-tune)
- **Method:** RAFT (Retrieval-Augmented Fine-Tuning), Unsloth LoRA
- **Epochs:** 4
- **Training hardware:** RTX 3090 Ti (homelab)
- **Dataset:** 4,535 curated examples (4,185 v2.0 base + 350 new)
- **Scope:** Rust + Dioxus 0.7.5 + Tailwind v4 + WCAG 2.2 AAA +
  fullstack server functions
- **Quantization:** GGUF Q4_K_M (9 GB) and MLX 4-bit (7.8 GB, 4.5 bpw)
- **Author:** Kevin Miller, Jr.

## Install via Ollama

```
ollama pull rockypod/neotoi-coder
```

## Read the Full Story

**[Read the whole story on RockyPod.com →](https://rockypod.com/blog/neotoi-coder-v2-release)**

---

## Files

| File | Format | Size | Use case |
|---|---|---|---|
| `neotoi-coder-v3-q4_k_m_patched.gguf` | GGUF Q4_K_M | 9 GB | LM Studio, llama.cpp, Ollama |
| `mlx-v3/` | MLX 4-bit (4.5 bpw) | 7.8 GB | Apple Silicon via mlx_lm / Ollama 0.19+ |
| `neotoi-coder-v2-q4_k_m.gguf` | GGUF Q4_K_M | 8.4 GB | v2.0 legacy |
| `mlx/` | MLX 4-bit | 7.5 GB | v2.0 legacy |

## Enabling Thinking Mode

### LM Studio

| Field | Value |
|---|---|
| Before System | `<\|im_start\|>system` |
| After System | `<\|im_end\|>` |
| Before User | `<\|im_start\|>user` |
| After User | `<\|im_end\|>` |
| Before Assistant | `<\|im_start\|>assistant\n<think>` |
| After Assistant | `<\|im_end\|>` |

### Ollama (GGUF)

```
FROM neotoi-coder-v3-q4_k_m_patched.gguf
PARAMETER temperature 0.2
PARAMETER num_ctx 16384
PARAMETER stop "<|im_end|>"
TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>
"""
SYSTEM You are Neotoi, an expert Rust and Dioxus 0.7 developer.
```

Or simply pull the published model:
```
ollama pull rockypod/neotoi-coder
```

### llama.cpp

```bash
./llama-cli \
  -m neotoi-coder-v3-q4_k_m_patched.gguf \
  -ngl 99 \
  --temp 0.2 \
  -p "<|im_start|>user\nYour question<|im_end|>\n<|im_start|>assistant\n<think>"
```

## What It Knows

Everything v2.0 knew, plus:

- Native scoped CSS via `css!()` macro (0.7.3)
- Native CSS modules with `.module.css` imports (0.7.3)
- `onauxclick` (middle-click) and `onscrollend` event handlers (0.7.3)
- Real WebSocket Stream+Sink — `stream.next()`, `sink.send()` (0.7.4)
- GlobalSignal cache rebuild patterns
- T11 server functions — `#[server]` with extractors, fullstack
  WebSocket one-liner, `ServerFnError` with HTTP status codes (0.7.3)
- `use_context_provider` / `use_context` placement — body only, never
  inside rsx!

Carried forward from v2.0: Dioxus 0.7 RSX brace syntax (never function-
call), `use_signal`, `use_resource` three-arm match, `r#for` on labels
only, `GlobalSignal` `.write()` semantics, WCAG 2.2 AAA (tooltip always
in DOM, listbox/option nesting, `aria_labelledby` on role containers),
dioxus-primitives discipline, `styles!()` macro, Tailwind v4 utilities
and semantic tokens, EN/VI i18n via pre-rsx! let bindings, dark mode
via `document::eval`, static content navigation with `use_memo`,
`use_context` panic behavior, `WritableResultExt`.

## Known Limitations

- **T2 RSX precision at Q14 / Q15 / Q22** — attribute placement
  regression vs v2.0; v3.1 target
- **Non-Dioxus web frameworks** — out of scope by design
- **Playwright / E2E testing** — out of scope (see the SvelteCoder line)

## Transparency

Full dataset, exam questions, and per-question model outputs are
published alongside the weights:

- **Weights:** [HuggingFace — rockypod/neotoi-coder](https://huggingface.co/rockypod/neotoi-coder)
- **Dataset + exam + per-question results:** [GitHub — rockypod/neotoi-coder](https://github.com/rockypod/neotoi-coder)
- **Ollama:** `ollama pull rockypod/neotoi-coder`

## License

Neotoi Coder Community License v1.0 — see LICENSE file.
Commercial use of model outputs permitted.
Weight redistribution prohibited.
Mental health deployment requires written permission.

## Credits

Built with:
- [Unsloth](https://github.com/unslothai/unsloth) — 2x faster fine-tuning
- [TRL](https://github.com/huggingface/trl) — SFTTrainer
- [Qwen3-Coder-14B](https://huggingface.co/Qwen/Qwen3-Coder-14B) — base model
- [MLX](https://github.com/ml-explore/mlx) — Apple Silicon inference
- [Claude Code](https://claude.ai/code) — dataset pipeline and training infrastructure
- [Dioxus](https://dioxuslabs.com) — the framework this model specializes in