Hugging Face's logo

rockypod
/
neotoi-coder

Text Generation
Safetensors
GGUF
Rust
English
Vietnamese
dioxus
accessibility
wcag
fine-tuned
raft
code
server-functions
qwen3
family-hub
scoped-css
syncstore
conversational
Model card Files
xet
 Community
1
rockypod commited on
Commit
dad7d0f
·
verified ·
1 Parent(s): 8b867da

Family hub README: rebrand 14B variant as 15B (matches Ollama tag); drop legacy framing — all three sizes are co-equal current variants

Browse files
Files changed (1) hide show
  1. README.md +29 -22
README.md CHANGED
@@ -20,20 +20,27 @@ tags:
20
  pipeline_tag: text-generation
21
  ---
22
 
23
- # Neotoi Coder — Family Hub
24
 
25
- A Rust / Dioxus 0.7 specialist LLM. v3.1 ships in **three sizes**, each as
26
- its own standalone repo:
 
 
 
 
 
 
 
27
 
28
  | Variant | Repo | Base | Params | Q4_K_M | Spec exam (104Q weighted, max 144.5) |
29
  |---|---|---|---|---|---|
30
  | **8B** (flagship) | [`rockypod/neotoi-coder-8b`](https://huggingface.co/rockypod/neotoi-coder-8b) | Qwen3-8B | 8.2B (6.95B non-embed) | 4.68 GB | **144.5 / 144.5 — 100.00%** |
31
  | 4B | [`rockypod/neotoi-coder-4b`](https://huggingface.co/rockypod/neotoi-coder-4b) | Qwen3-4B | 4.0B (3.6B non-embed, tied) | 2.33 GB | 143.5 / 144.5 — 99.31% |
32
- | 14B (legacy) | this repo (`rockypod/neotoi-coder`) | Qwen3-Coder-14B | 14.8B (13.2B non-embed) | 8.40 GB | 137.0 / 144.5 — 94.81% |
33
 
34
- All three clear the 90% publication bar **and** the 95% release bar with all per-tier floors PASS. The **8B is the recommended default**; pick the **4B** if disk / RAM is tight (or for ~40% faster generation), pick the **14B** for the broadest coverage of legacy material.
35
 
36
- > Each variant lives in its **own model repo** so it's separately searchable and discoverable on HuggingFace. This page (`rockypod/neotoi-coder`) is the family hub *and* still hosts the legacy 14B GGUFs.
37
 
38
  ## Install via Ollama
39
 
@@ -44,7 +51,7 @@ ollama pull rockypod/neotoi-coder:8b
44
  # 4B — disk / RAM constrained, ~40% faster generation
45
  ollama pull rockypod/neotoi-coder:4b
46
 
47
- # 14B — legacy, broadest coverage
48
  ollama pull rockypod/neotoi-coder:15b
49
  ```
50
 
@@ -52,7 +59,7 @@ ollama pull rockypod/neotoi-coder:15b
52
 
53
  Re-graded 2026-04-26 with the patched `run_grade_v31.py` (Q87 now also accepts `LANG()` / `THEME()` GlobalSignal accessor calls in addition to the literal `Signal` token — a false-negative fix that unlocked the 8B's perfect score).
54
 
55
- | Tier | Max wt | 8B | 4B | 14B |
56
  |---|---|---|---|---|
57
  | T1 Fundamentals | 12.0 | 12.0 ✅ | 11.0 ⚠️ 91.7% | 12.0 ✅ |
58
  | T2 RSX Syntax | 12.0 | 12.0 ✅ | 12.0 ✅ | 10.0 ⚠️ 83.3% |
@@ -73,12 +80,12 @@ Tier floors (82% on weight-1.0 / 1.5 tiers, 88% on weight-2.0 tiers): all PASS f
73
 
74
  ## What's new in v3.1 (vs v3.0)
75
 
76
- - **Two new sizes**: 8B and 4B alongside the 14B base, both surpassing the 14B's score.
77
- - **T1 Fundamentals → 100%** on 8B and 14B, 91.7% on 4B (+8.3 pts vs v3.0 14B).
78
- - **T6 Hard Reasoning → 100%** clean sweep, all three variants (+25 pts vs v3.0 14B).
79
  - **T8 GlobalSignal / i18n → 100%** all three variants.
80
  - **T10 Dioxus 0.7.4 → 100%** all three variants.
81
- - **8 tiers at 100%** on the 14B; **11 tiers at 100%** on the 8B (perfect).
82
  - **Dataset:** 4,880 curated examples across 43 topics (up from 4,535).
83
 
84
  ## Version History
@@ -88,20 +95,20 @@ Tier floors (82% on weight-1.0 / 1.5 tiers, 88% on weight-2.0 tiers): all PASS f
88
  | v1.0 | Qwen3-Coder-14B (14.8B) | 51/60 (85.0%) | 60Q standard | — |
89
  | v2.0 | Qwen3-Coder-14B (14.8B) | 135.5/140 (96.8%) | 100Q weighted | 4,185 |
90
  | v3.0 | Qwen3-Coder-14B (14.8B) | 124.0/144.5 (85.8%) | 103Q weighted | 4,535 |
91
- | v3.1 14B | Qwen3-Coder-14B (14.8B) | 137.0/144.5 (94.81%) | 103Q weighted | 4,880 |
92
  | **v3.1 8B** | **Qwen3-8B (8.2B)** | **144.5/144.5 (100.00%)** | **103Q weighted** | **4,880** |
93
  | v3.1 4B | Qwen3-4B (4.0B, tied) | 143.5/144.5 (99.31%) | 103Q weighted | 4,880 |
94
 
95
- ## Files in this repo (`rockypod/neotoi-coder`, 14B legacy GGUFs)
96
 
97
  | File | Format | Size | Use case |
98
  |---|---|---|---|
99
- | `neotoi-coder-v3.1-q4_k_m.gguf` | GGUF Q4_K_M | 8.4 GB | LM Studio, llama.cpp, Ollama (current 14B) |
100
- | `neotoi-coder-v3-q4_k_m_patched.gguf` | GGUF Q4_K_M | 9 GB | v3.0 legacy |
101
- | `neotoi-coder-v2.0-q4_k_m.gguf` | GGUF Q4_K_M | 9 GB | v2.0 legacy |
102
- | `neotoi-coder-v1-q4_k_m_final.gguf` | GGUF Q4_K_M | 9 GB | v1.0 legacy |
103
 
104
- For the **8B** and **4B** Q4_K_M GGUFs, go to their dedicated repos:
105
 
106
  - https://huggingface.co/rockypod/neotoi-coder-8b
107
  - https://huggingface.co/rockypod/neotoi-coder-4b
@@ -121,7 +128,7 @@ This model emits Qwen3 native `<think>...</think>` blocks. Thinking is on by def
121
  | Before Assistant | `<\|im_start\|>assistant\n<think>` |
122
  | After Assistant | `<\|im_end\|>` |
123
 
124
- ### Ollama (custom Modelfile, 14B)
125
 
126
  ```Modelfile
127
  FROM neotoi-coder-v3.1-q4_k_m.gguf
@@ -170,13 +177,13 @@ ollama pull rockypod/neotoi-coder:15b
170
 
171
  ## Known Limitations
172
 
173
- - **rsx! macro drops on the 14B** for 6 RSX-heavy questions (Q17 / 22 / 30 / 37 / 39 / 43); v3.2 target. The 8B and 4B do not reproduce these misses.
174
  - **Non-Dioxus web frameworks** — out of scope by design (SvelteKit coverage lives in `rockypod/svcoder`).
175
  - **Playwright / E2E testing** — out of scope.
176
 
177
  ## Transparency
178
 
179
- - **Per-variant weights:** [`-8b`](https://huggingface.co/rockypod/neotoi-coder-8b) · [`-4b`](https://huggingface.co/rockypod/neotoi-coder-4b) · this repo (14B)
180
  - **Exam runner, grader, per-question results:** [GitHub — rockypod/neotoi-coder](https://github.com/rockypod/neotoi-coder)
181
  - **Ollama:** `ollama pull rockypod/neotoi-coder:8b` (or `:4b`, or `:15b`)
182
 
 
20
  pipeline_tag: text-generation
21
  ---
22
 
23
+ # Neotoi Coder
24
 
25
+ A Rust / Dioxus 0.7 specialist LLM. v3.1 ships in **three sizes** — 15B, 8B,
26
+ and 4B — all fine-tuned via RAFT (Retrieval-Augmented Fine-Tuning) on Qwen3
27
+ base models. Optimized for production-quality Dioxus 0.7 components with
28
+ Tailwind v4 and WCAG 2.2 AAA accessibility.
29
+
30
+ All three are current. They were trained from the same v3.1 dataset, exam
31
+ the same way, and ship together. Pick based on hardware, not currency.
32
+
33
+ ## Variants
34
 
35
  | Variant | Repo | Base | Params | Q4_K_M | Spec exam (104Q weighted, max 144.5) |
36
  |---|---|---|---|---|---|
37
  | **8B** (flagship) | [`rockypod/neotoi-coder-8b`](https://huggingface.co/rockypod/neotoi-coder-8b) | Qwen3-8B | 8.2B (6.95B non-embed) | 4.68 GB | **144.5 / 144.5 — 100.00%** |
38
  | 4B | [`rockypod/neotoi-coder-4b`](https://huggingface.co/rockypod/neotoi-coder-4b) | Qwen3-4B | 4.0B (3.6B non-embed, tied) | 2.33 GB | 143.5 / 144.5 — 99.31% |
39
+ | 15B | this repo (`rockypod/neotoi-coder`) | Qwen3-Coder-14B | 14.8B (13.2B non-embed) | 8.40 GB | 137.0 / 144.5 — 94.81% |
40
 
41
+ All three clear the 90% publication bar **and** the 95% release bar with all per-tier floors PASS. The **8B is the recommended default**; pick the **4B** if disk / RAM is tight (or for ~40% faster generation), pick the **15B** for the broadest coverage and the most context-rich generations.
42
 
43
+ > Each variant lives in its **own model repo** for searchability. This page (`rockypod/neotoi-coder`) is the family hub *and* hosts the 15B GGUFs.
44
 
45
  ## Install via Ollama
46
 
 
51
  # 4B — disk / RAM constrained, ~40% faster generation
52
  ollama pull rockypod/neotoi-coder:4b
53
 
54
+ # 15B — largest, broadest coverage
55
  ollama pull rockypod/neotoi-coder:15b
56
  ```
57
 
 
59
 
60
  Re-graded 2026-04-26 with the patched `run_grade_v31.py` (Q87 now also accepts `LANG()` / `THEME()` GlobalSignal accessor calls in addition to the literal `Signal` token — a false-negative fix that unlocked the 8B's perfect score).
61
 
62
+ | Tier | Max wt | 8B | 4B | 15B |
63
  |---|---|---|---|---|
64
  | T1 Fundamentals | 12.0 | 12.0 ✅ | 11.0 ⚠️ 91.7% | 12.0 ✅ |
65
  | T2 RSX Syntax | 12.0 | 12.0 ✅ | 12.0 ✅ | 10.0 ⚠️ 83.3% |
 
80
 
81
  ## What's new in v3.1 (vs v3.0)
82
 
83
+ - **Three sizes**: 8B and 4B alongside the 15B base, both surpassing the 15B's score.
84
+ - **T1 Fundamentals → 100%** on 8B and 15B, 91.7% on 4B (+8.3 pts vs v3.0).
85
+ - **T6 Hard Reasoning → 100%** clean sweep, all three variants (+25 pts vs v3.0).
86
  - **T8 GlobalSignal / i18n → 100%** all three variants.
87
  - **T10 Dioxus 0.7.4 → 100%** all three variants.
88
+ - **8 tiers at 100%** on the 15B; **11 tiers at 100%** on the 8B (perfect).
89
  - **Dataset:** 4,880 curated examples across 43 topics (up from 4,535).
90
 
91
  ## Version History
 
95
  | v1.0 | Qwen3-Coder-14B (14.8B) | 51/60 (85.0%) | 60Q standard | — |
96
  | v2.0 | Qwen3-Coder-14B (14.8B) | 135.5/140 (96.8%) | 100Q weighted | 4,185 |
97
  | v3.0 | Qwen3-Coder-14B (14.8B) | 124.0/144.5 (85.8%) | 103Q weighted | 4,535 |
98
+ | v3.1 15B | Qwen3-Coder-14B (14.8B) | 137.0/144.5 (94.81%) | 103Q weighted | 4,880 |
99
  | **v3.1 8B** | **Qwen3-8B (8.2B)** | **144.5/144.5 (100.00%)** | **103Q weighted** | **4,880** |
100
  | v3.1 4B | Qwen3-4B (4.0B, tied) | 143.5/144.5 (99.31%) | 103Q weighted | 4,880 |
101
 
102
+ ## Files in this repo (`rockypod/neotoi-coder`, 15B and historical)
103
 
104
  | File | Format | Size | Use case |
105
  |---|---|---|---|
106
+ | `neotoi-coder-v3.1-q4_k_m.gguf` | GGUF Q4_K_M | 8.4 GB | LM Studio, llama.cpp, Ollama (current 15B) |
107
+ | `neotoi-coder-v3-q4_k_m_patched.gguf` | GGUF Q4_K_M | 9 GB | v3.0 archive |
108
+ | `neotoi-coder-v2.0-q4_k_m.gguf` | GGUF Q4_K_M | 9 GB | v2.0 archive |
109
+ | `neotoi-coder-v1-q4_k_m_final.gguf` | GGUF Q4_K_M | 9 GB | v1.0 archive |
110
 
111
+ For the **8B** and **4B** Q4_K_M GGUFs, see their dedicated repos:
112
 
113
  - https://huggingface.co/rockypod/neotoi-coder-8b
114
  - https://huggingface.co/rockypod/neotoi-coder-4b
 
128
  | Before Assistant | `<\|im_start\|>assistant\n<think>` |
129
  | After Assistant | `<\|im_end\|>` |
130
 
131
+ ### Ollama (custom Modelfile, 15B)
132
 
133
  ```Modelfile
134
  FROM neotoi-coder-v3.1-q4_k_m.gguf
 
177
 
178
  ## Known Limitations
179
 
180
+ - **rsx! macro drops on the 15B** for 6 RSX-heavy questions (Q17 / 22 / 30 / 37 / 39 / 43); v3.2 target. The 8B and 4B do not reproduce these misses.
181
  - **Non-Dioxus web frameworks** — out of scope by design (SvelteKit coverage lives in `rockypod/svcoder`).
182
  - **Playwright / E2E testing** — out of scope.
183
 
184
  ## Transparency
185
 
186
+ - **Per-variant weights:** [`-8b`](https://huggingface.co/rockypod/neotoi-coder-8b) · [`-4b`](https://huggingface.co/rockypod/neotoi-coder-4b) · this repo (15B)
187
  - **Exam runner, grader, per-question results:** [GitHub — rockypod/neotoi-coder](https://github.com/rockypod/neotoi-coder)
188
  - **Ollama:** `ollama pull rockypod/neotoi-coder:8b` (or `:4b`, or `:15b`)
189