guell00 commited on
Commit
bd6894d
·
verified ·
1 Parent(s): 3df108f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +1030 -16
README.md CHANGED
@@ -1,25 +1,1039 @@
1
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  tags:
 
 
 
 
 
3
  - gguf
4
- - llama.cpp
5
  - unsloth
6
- - vision-language-model
 
 
 
 
 
 
 
 
7
  ---
8
 
9
- # OBSIDIAN-9B-Coder : GGUF
10
 
11
- This model was finetuned and converted to GGUF format using [Unsloth](https://github.com/unslothai/unsloth).
 
 
 
 
 
 
12
 
13
- **Example usage**:
14
- - For text only LLMs: `llama-cli -hf guell00/OBSIDIAN-9B-Coder --jinja`
15
- - For multimodal models: `llama-mtmd-cli -hf guell00/OBSIDIAN-9B-Coder --jinja`
 
 
16
 
17
- ## Available Model files:
18
- - `Qwopus3.5-9B-Coder.Q3_K_M.gguf`
19
- - `Qwopus3.5-9B-Coder.Q8_0.gguf`
20
- - `Qwopus3.5-9B-Coder.Q6_K.gguf`
21
- - `Qwopus3.5-9B-Coder.Q5_K_M.gguf`
22
- - `Qwopus3.5-9B-Coder.Q4_K_M.gguf`
23
- - `Qwopus3.5-9B-Coder.BF16-mmproj.gguf`
24
- This was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)
25
- [<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ base_model: Jackrong/Qwopus3.5-9B-Coder
3
+ datasets:
4
+ - guell00/Coder-max
5
+ language:
6
+ - pt
7
+ - en
8
+ - es
9
+ - fr
10
+ - de
11
+ - it
12
+ - ja
13
+ - zh
14
+ - ru
15
+ - tr
16
+ pipeline_tag: text-generation
17
  tags:
18
+ - code
19
+ - coding
20
+ - coder
21
+ - qwen
22
+ - qwen3.5
23
  - gguf
 
24
  - unsloth
25
+ - lora
26
+ - sft
27
+ - html
28
+ - css
29
+ - javascript
30
+ - threejs
31
+ - canvas
32
+ - python
33
+ - multilingual
34
  ---
35
 
36
+ <div align="center">
37
 
38
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 360" width="100%" height="auto">
39
+ <defs>
40
+ <linearGradient id="bg" x1="0%" y1="0%" x2="100%" y2="100%">
41
+ <stop offset="0%" stop-color="#050507"/>
42
+ <stop offset="45%" stop-color="#0d0914"/>
43
+ <stop offset="100%" stop-color="#071116"/>
44
+ </linearGradient>
45
 
46
+ <linearGradient id="obsidian" x1="0%" y1="0%" x2="100%" y2="0%">
47
+ <stop offset="0%" stop-color="#7c3aed"/>
48
+ <stop offset="45%" stop-color="#c026d3"/>
49
+ <stop offset="100%" stop-color="#06b6d4"/>
50
+ </linearGradient>
51
 
52
+ <filter id="glow">
53
+ <feGaussianBlur stdDeviation="5" result="blur"/>
54
+ <feMerge>
55
+ <feMergeNode in="blur"/>
56
+ <feMergeNode in="SourceGraphic"/>
57
+ </feMerge>
58
+ </filter>
59
+ </defs>
60
+
61
+ <rect width="1200" height="360" rx="22" fill="url(#bg)"/>
62
+
63
+ <rect x="2" y="2" width="1196" height="356" rx="20"
64
+ fill="none"
65
+ stroke="url(#obsidian)"
66
+ stroke-width="2"
67
+ opacity="0.5"/>
68
+
69
+ <g opacity="0.07" stroke="#a855f7">
70
+ <path d="M0 90H1200 M0 180H1200 M0 270H1200"/>
71
+ <path d="M150 0V360 M300 0V360 M450 0V360 M600 0V360 M750 0V360 M900 0V360 M1050 0V360"/>
72
+ </g>
73
+
74
+ <g transform="translate(85,70)" filter="url(#glow)">
75
+ <path d="M100 0 L180 62 L155 185 L100 225 L45 185 L20 62 Z"
76
+ fill="#0c0912"
77
+ stroke="url(#obsidian)"
78
+ stroke-width="3"/>
79
+
80
+ <path d="M100 25 L145 75 L128 160 L100 195 L72 160 L55 75 Z"
81
+ fill="#171020"
82
+ stroke="#8b5cf6"
83
+ stroke-width="2"/>
84
+
85
+ <path d="M100 45 L120 100 L100 170 L80 100 Z"
86
+ fill="url(#obsidian)"/>
87
+ </g>
88
+
89
+ <text x="330" y="125"
90
+ font-family="-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif"
91
+ font-size="58"
92
+ font-weight="900"
93
+ fill="url(#obsidian)"
94
+ letter-spacing="3">
95
+ OBSIDIAN
96
+ </text>
97
+
98
+ <text x="334" y="177"
99
+ font-family="-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif"
100
+ font-size="30"
101
+ font-weight="800"
102
+ fill="#e2e8f0"
103
+ letter-spacing="6">
104
+ 9B CODER
105
+ </text>
106
+
107
+ <text x="334" y="218"
108
+ font-family="monospace"
109
+ font-size="16"
110
+ fill="#94a3b8">
111
+ QWEN CODER • MULTILINGUAL • FRONTEND SPECIALIZED
112
+ </text>
113
+
114
+ <g transform="translate(334,250)">
115
+ <rect width="130" height="30" rx="15"
116
+ fill="#7c3aed"
117
+ fill-opacity="0.18"/>
118
+ <text x="65" y="20"
119
+ text-anchor="middle"
120
+ font-family="monospace"
121
+ font-size="12"
122
+ fill="#c4b5fd">
123
+ THREE.JS
124
+ </text>
125
+
126
+ <rect x="145" width="130" height="30" rx="15"
127
+ fill="#c026d3"
128
+ fill-opacity="0.16"/>
129
+ <text x="210" y="20"
130
+ text-anchor="middle"
131
+ font-family="monospace"
132
+ font-size="12"
133
+ fill="#f0abfc">
134
+ JAVASCRIPT
135
+ </text>
136
+
137
+ <rect x="290" width="130" height="30" rx="15"
138
+ fill="#2563eb"
139
+ fill-opacity="0.16"/>
140
+ <text x="355" y="20"
141
+ text-anchor="middle"
142
+ font-family="monospace"
143
+ font-size="12"
144
+ fill="#93c5fd">
145
+ HTML/CSS
146
+ </text>
147
+
148
+ <rect x="435" width="130" height="30" rx="15"
149
+ fill="#0891b2"
150
+ fill-opacity="0.16"/>
151
+ <text x="500" y="20"
152
+ text-anchor="middle"
153
+ font-family="monospace"
154
+ font-size="12"
155
+ fill="#67e8f9">
156
+ CANVAS
157
+ </text>
158
+ </g>
159
+
160
+ <text x="334" y="320"
161
+ font-family="monospace"
162
+ font-size="13"
163
+ fill="#64748b">
164
+ FORGED FOR COMPLETE CODE GENERATION
165
+ </text>
166
+ </svg>
167
+
168
+ </div>
169
+
170
+ # OBSIDIAN-9B-Coder
171
+
172
+ **OBSIDIAN-9B-Coder** is a 9B coding model fine-tuned on the **Coder Max Multilingual** dataset.
173
+
174
+ The model is specialized in generating **complete software implementations**, with a strong focus on modern frontend development, interactive browser applications, Three.js, HTML5 Canvas, JavaScript, HTML/CSS and general programming.
175
+
176
+ > **Complete code. Long context. Interactive software.**
177
+
178
+ ---
179
+
180
+ <div align="center">
181
+
182
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 125" width="100%">
183
+ <rect width="1200" height="125" rx="18"
184
+ fill="#08080d"
185
+ stroke="#312e81"
186
+ stroke-opacity="0.7"/>
187
+
188
+ <g font-family="monospace" text-anchor="middle">
189
+
190
+ <text x="120" y="48"
191
+ font-size="27"
192
+ font-weight="900"
193
+ fill="#c4b5fd">9B</text>
194
+ <text x="120" y="78"
195
+ font-size="11"
196
+ fill="#64748b">MODEL CLASS</text>
197
+
198
+ <text x="360" y="48"
199
+ font-size="27"
200
+ font-weight="900"
201
+ fill="#67e8f9">32K</text>
202
+ <text x="360" y="78"
203
+ font-size="11"
204
+ fill="#64748b">TRAINING CONTEXT</text>
205
+
206
+ <text x="600" y="48"
207
+ font-size="27"
208
+ font-weight="900"
209
+ fill="#f0abfc">10</text>
210
+ <text x="600" y="78"
211
+ font-size="11"
212
+ fill="#64748b">LANGUAGES</text>
213
+
214
+ <text x="840" y="48"
215
+ font-size="27"
216
+ font-weight="900"
217
+ fill="#93c5fd">LoRA</text>
218
+ <text x="840" y="78"
219
+ font-size="11"
220
+ fill="#64748b">SFT</text>
221
+
222
+ <text x="1080" y="48"
223
+ font-size="27"
224
+ font-weight="900"
225
+ fill="#86efac">GGUF</text>
226
+ <text x="1080" y="78"
227
+ font-size="11"
228
+ fill="#64748b">READY</text>
229
+
230
+ </g>
231
+ </svg>
232
+
233
+ </div>
234
+
235
+ ## Overview
236
+
237
+ OBSIDIAN-9B-Coder was created to specialize a Qwen-based coding model toward **implementation-heavy programming tasks**.
238
+
239
+ Instead of focusing primarily on explanations surrounding code, the fine-tuning corpus heavily emphasizes the generation of the actual implementation.
240
+
241
+ The model is particularly suited for:
242
+
243
+ - Complete single-file web applications
244
+ - HTML5 and modern CSS
245
+ - JavaScript ES6+
246
+ - Three.js
247
+ - HTML5 Canvas
248
+ - Interactive browser applications
249
+ - Games and graphical experiments
250
+ - DOM manipulation
251
+ - Animation loops
252
+ - State management
253
+ - Python
254
+ - Algorithmic programming
255
+ - Multilingual programming instructions
256
+
257
+ ---
258
+
259
+ ## Core Philosophy
260
+
261
+ <div align="center">
262
+
263
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 220" width="100%">
264
+ <defs>
265
+ <linearGradient id="flow" x1="0%" y1="0%" x2="100%" y2="0%">
266
+ <stop offset="0%" stop-color="#7c3aed"/>
267
+ <stop offset="50%" stop-color="#c026d3"/>
268
+ <stop offset="100%" stop-color="#0891b2"/>
269
+ </linearGradient>
270
+ </defs>
271
+
272
+ <rect width="1200" height="220" rx="18"
273
+ fill="#07070b"
274
+ stroke="#27272a"/>
275
+
276
+ <text x="600" y="40"
277
+ text-anchor="middle"
278
+ font-family="monospace"
279
+ font-size="15"
280
+ font-weight="700"
281
+ fill="#c4b5fd">
282
+ FROM PROMPT TO COMPLETE SOFTWARE
283
+ </text>
284
+
285
+ <g font-family="monospace" text-anchor="middle">
286
+
287
+ <rect x="75" y="80" width="230" height="70" rx="14"
288
+ fill="#120d1b"
289
+ stroke="#7c3aed"/>
290
+
291
+ <text x="190" y="109"
292
+ font-size="14"
293
+ fill="#ddd6fe">
294
+ USER REQUEST
295
+ </text>
296
+
297
+ <text x="190" y="132"
298
+ font-size="11"
299
+ fill="#64748b">
300
+ natural language
301
+ </text>
302
+
303
+ <path d="M305 115H445"
304
+ stroke="url(#flow)"
305
+ stroke-width="4"/>
306
+
307
+ <path d="M432 105L445 115L432 125"
308
+ fill="none"
309
+ stroke="#c026d3"
310
+ stroke-width="4"/>
311
+
312
+ <rect x="445" y="80" width="310" height="70" rx="14"
313
+ fill="#150d18"
314
+ stroke="#c026d3"/>
315
+
316
+ <text x="600" y="109"
317
+ font-size="14"
318
+ fill="#f5d0fe">
319
+ OBSIDIAN-9B-CODER
320
+ </text>
321
+
322
+ <text x="600" y="132"
323
+ font-size="11"
324
+ fill="#64748b">
325
+ generate + construct
326
+ </text>
327
+
328
+ <path d="M755 115H895"
329
+ stroke="url(#flow)"
330
+ stroke-width="4"/>
331
+
332
+ <path d="M882 105L895 115L882 125"
333
+ fill="none"
334
+ stroke="#0891b2"
335
+ stroke-width="4"/>
336
+
337
+ <rect x="895" y="80" width="230" height="70" rx="14"
338
+ fill="#07171b"
339
+ stroke="#0891b2"/>
340
+
341
+ <text x="1010" y="109"
342
+ font-size="14"
343
+ fill="#a5f3fc">
344
+ COMPLETE CODE
345
+ </text>
346
+
347
+ <text x="1010" y="132"
348
+ font-size="11"
349
+ fill="#64748b">
350
+ full implementation
351
+ </text>
352
+
353
+ </g>
354
+
355
+ <text x="600" y="190"
356
+ text-anchor="middle"
357
+ font-family="monospace"
358
+ font-size="11"
359
+ fill="#52525b">
360
+ LESS BOILERPLATE • MORE IMPLEMENTATION • COMPLETE OUTPUTS
361
+ </text>
362
+ </svg>
363
+
364
+ </div>
365
+
366
+ ---
367
+
368
+ # GGUF
369
+
370
+ This model was fine-tuned and converted to **GGUF** format using **Unsloth**.
371
+
372
+ The repository contains multiple quantizations so users can select the appropriate balance between model size, memory usage and output quality.
373
+
374
+ ## Available Model Files
375
+
376
+ | File | Quantization | Notes |
377
+ |---|---|---|
378
+ | `Qwopus3.5-9B-Coder.Q8_0.gguf` | Q8_0 | Highest-quality quantized version |
379
+ | `Qwopus3.5-9B-Coder.Q6_K.gguf` | Q6_K | High quality with reduced size |
380
+ | `Qwopus3.5-9B-Coder.Q5_K_M.gguf` | Q5_K_M | Strong quality/size balance |
381
+ | `Qwopus3.5-9B-Coder.Q4_K_M.gguf` | Q4_K_M | Recommended balance for local use |
382
+ | `Qwopus3.5-9B-Coder.Q3_K_M.gguf` | Q3_K_M | Lower memory requirement |
383
+ | `Qwopus3.5-9B-Coder.BF16-mmproj.gguf` | BF16 mmproj | Multimodal projector artifact |
384
+
385
+ ---
386
+
387
+ ## llama.cpp
388
+
389
+ For compatible text-generation configurations:
390
+
391
+ ```bash
392
+ llama-cli -hf guell00/OBSIDIAN-9B-Coder --jinja
393
+ ```
394
+
395
+ For compatible multimodal configurations:
396
+
397
+ ```bash
398
+ llama-mtmd-cli -hf guell00/OBSIDIAN-9B-Coder --jinja
399
+ ```
400
+
401
+ Depending on your environment and the files selected automatically by `llama.cpp`, you may want to explicitly choose a GGUF quantization.
402
+
403
+ ---
404
+
405
+ ## Quantization Guide
406
+
407
+ <div align="center">
408
+
409
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 310" width="100%">
410
+ <rect width="1200" height="310" rx="20"
411
+ fill="#08080d"
412
+ stroke="#27272a"/>
413
+
414
+ <text x="600" y="42"
415
+ text-anchor="middle"
416
+ font-family="monospace"
417
+ font-size="16"
418
+ font-weight="700"
419
+ fill="#c4b5fd">
420
+ GGUF QUANTIZATION LADDER
421
+ </text>
422
+
423
+ <g font-family="monospace">
424
+
425
+ <text x="90" y="92" fill="#94a3b8">Q8_0</text>
426
+ <rect x="200" y="75" width="780" height="20" rx="10" fill="#18181b"/>
427
+ <rect x="200" y="75" width="760" height="20" rx="10" fill="#7c3aed"/>
428
+ <text x="1010" y="91" fill="#c4b5fd">QUALITY</text>
429
+
430
+ <text x="90" y="132" fill="#94a3b8">Q6_K</text>
431
+ <rect x="200" y="115" width="780" height="20" rx="10" fill="#18181b"/>
432
+ <rect x="200" y="115" width="700" height="20" rx="10" fill="#8b5cf6"/>
433
+ <text x="1010" y="131" fill="#c4b5fd">HIGH</text>
434
+
435
+ <text x="90" y="172" fill="#94a3b8">Q5_K_M</text>
436
+ <rect x="200" y="155" width="780" height="20" rx="10" fill="#18181b"/>
437
+ <rect x="200" y="155" width="640" height="20" rx="10" fill="#c026d3"/>
438
+ <text x="1010" y="171" fill="#f0abfc">BALANCED+</text>
439
+
440
+ <text x="90" y="212" fill="#94a3b8">Q4_K_M</text>
441
+ <rect x="200" y="195" width="780" height="20" rx="10" fill="#18181b"/>
442
+ <rect x="200" y="195" width="570" height="20" rx="10" fill="#2563eb"/>
443
+ <text x="1010" y="211" fill="#93c5fd">BALANCED</text>
444
+
445
+ <text x="90" y="252" fill="#94a3b8">Q3_K_M</text>
446
+ <rect x="200" y="235" width="780" height="20" rx="10" fill="#18181b"/>
447
+ <rect x="200" y="235" width="470" height="20" rx="10" fill="#0891b2"/>
448
+ <text x="1010" y="251" fill="#67e8f9">COMPACT</text>
449
+
450
+ </g>
451
+
452
+ <text x="600" y="288"
453
+ text-anchor="middle"
454
+ font-family="monospace"
455
+ font-size="11"
456
+ fill="#52525b">
457
+ MORE BITS → MORE MEMORY • FEWER BITS → SMALLER FOOTPRINT
458
+ </text>
459
+ </svg>
460
+
461
+ </div>
462
+
463
+ For most local deployments, **Q4_K_M** or **Q5_K_M** are reasonable starting points.
464
+
465
+ Use **Q8_0** when memory is less constrained and preserving model quality is more important.
466
+
467
+ ---
468
+
469
+ # Training Dataset
470
+
471
+ OBSIDIAN-9B-Coder was fine-tuned using:
472
+
473
+ **Coder Max Multilingual**
474
+ `guell00/Coder-max`
475
+
476
+ Coder Max is a conversational supervised fine-tuning dataset focused heavily on code generation.
477
+
478
+ The corpus was designed to contain complete implementations instead of heavily truncated programming answers.
479
+
480
+ ## Dataset Characteristics
481
+
482
+ | Characteristic | Description |
483
+ |---|---|
484
+ | Format | JSONL |
485
+ | Structure | Conversational messages |
486
+ | Training type | Supervised Fine-Tuning |
487
+ | Languages | 10 |
488
+ | Main content | Programming |
489
+ | Code density | ~95%+ |
490
+ | Frontend focus | Strong |
491
+ | Three.js specialization | Strong |
492
+ | Long code outputs | Preserved |
493
+
494
+ ---
495
+
496
+ ## Dataset Scale
497
+
498
+ Coder Max is distributed in incremental variants.
499
+
500
+ | Variant | Physical Size | Records | Messages | Code Density |
501
+ |---|---:|---:|---:|---:|
502
+ | `001MB` | 3,739,874 B | 109 | 220 | 99.08% |
503
+ | `010MB` | 12,709,008 B | 969 | 2,012 | 96.18% |
504
+ | `100MB` | 102,679,666 B | 9,790 | 20,396 | 95.84% |
505
+ | `300MB` | 302,689,973 B | 29,233 | 60,916 | 95.83% |
506
+ | `500MB` | 502,678,782 B | 48,676 | 101,442 | 95.82% |
507
+ | `600MB` | 602,666,385 B | 58,466 | 121,848 | 95.82% |
508
+ | `001GB` | 1,002,677,454 B | 97,499 | 203,200 | 95.82% |
509
+ | `total_4GB` | 4,002,669,404 B | 390,302 | 813,452 | 95.81% |
510
+
511
+ The larger variants contain the content represented by the smaller variants, allowing different training scales without requiring manual concatenation.
512
+
513
+ ---
514
+
515
+ # Specialization
516
+
517
+ <div align="center">
518
+
519
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 335" width="100%">
520
+ <rect width="1200" height="335" rx="20"
521
+ fill="#08080d"
522
+ stroke="#27272a"/>
523
+
524
+ <text x="600" y="43"
525
+ text-anchor="middle"
526
+ font-family="monospace"
527
+ font-size="16"
528
+ font-weight="700"
529
+ fill="#e2e8f0">
530
+ OBSIDIAN SPECIALIZATION
531
+ </text>
532
+
533
+ <g font-family="monospace" font-size="13">
534
+
535
+ <text x="85" y="95" fill="#94a3b8">Three.js</text>
536
+ <rect x="245" y="78" width="780" height="22" rx="11" fill="#18181b"/>
537
+ <rect x="245" y="78" width="755" height="22" rx="11" fill="#7c3aed"/>
538
+ <text x="1060" y="95" fill="#c4b5fd">CORE</text>
539
+
540
+ <text x="85" y="140" fill="#94a3b8">JavaScript</text>
541
+ <rect x="245" y="123" width="780" height="22" rx="11" fill="#18181b"/>
542
+ <rect x="245" y="123" width="740" height="22" rx="11" fill="#a21caf"/>
543
+ <text x="1060" y="140" fill="#f0abfc">HIGH</text>
544
+
545
+ <text x="85" y="185" fill="#94a3b8">HTML / CSS</text>
546
+ <rect x="245" y="168" width="780" height="22" rx="11" fill="#18181b"/>
547
+ <rect x="245" y="168" width="720" height="22" rx="11" fill="#2563eb"/>
548
+ <text x="1060" y="185" fill="#93c5fd">HIGH</text>
549
+
550
+ <text x="85" y="230" fill="#94a3b8">Canvas</text>
551
+ <rect x="245" y="213" width="780" height="22" rx="11" fill="#18181b"/>
552
+ <rect x="245" y="213" width="680" height="22" rx="11" fill="#0891b2"/>
553
+ <text x="1060" y="230" fill="#67e8f9">HIGH</text>
554
+
555
+ <text x="85" y="275" fill="#94a3b8">Python</text>
556
+ <rect x="245" y="258" width="780" height="22" rx="11" fill="#18181b"/>
557
+ <rect x="245" y="258" width="610" height="22" rx="11" fill="#4f46e5"/>
558
+ <text x="1060" y="275" fill="#a5b4fc">STRONG</text>
559
+
560
+ </g>
561
+ </svg>
562
+
563
+ </div>
564
+
565
+ ## Three.js
566
+
567
+ Three.js is one of the primary specialization targets of OBSIDIAN.
568
+
569
+ Training examples include patterns involving:
570
+
571
+ - Scene creation
572
+ - Perspective cameras
573
+ - Lighting
574
+ - Meshes
575
+ - Materials
576
+ - Geometry
577
+ - Animation loops
578
+ - Browser interaction
579
+ - Keyboard input
580
+ - Game mechanics
581
+ - 3D environments
582
+ - Interactive simulations
583
+ - Single-file applications
584
+
585
+ The goal is to teach the model to connect all required pieces of a Three.js application rather than merely outputting isolated API examples.
586
+
587
+ ---
588
+
589
+ ## Frontend Generation
590
+
591
+ The model has strong exposure to complete frontend applications combining:
592
+
593
+ ```text
594
+ HTML
595
+
596
+ ├── CSS
597
+
598
+ └── JavaScript
599
+
600
+ ├── DOM
601
+ ├── State
602
+ ├── Events
603
+ ├── Canvas
604
+ ├── Three.js
605
+ └── Animation
606
+ ```
607
+
608
+ A typical training target may contain an entire application:
609
+
610
+ ```html
611
+ <!DOCTYPE html>
612
+ <html>
613
+ <head>
614
+ <style>
615
+ /* Complete interface */
616
+ </style>
617
+ </head>
618
+
619
+ <body>
620
+ <main>
621
+ <!-- Complete application -->
622
+ </main>
623
+
624
+ <script>
625
+ // State
626
+ // Logic
627
+ // Rendering
628
+ // Interaction
629
+ // Animation loop
630
+ </script>
631
+ </body>
632
+ </html>
633
+ ```
634
+
635
+ ---
636
+
637
+ # Multilingual Programming
638
+
639
+ <div align="center">
640
+
641
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 180" width="100%">
642
+ <rect width="1200" height="180" rx="18"
643
+ fill="#09090e"
644
+ stroke="#27272a"/>
645
+
646
+ <text x="600" y="38"
647
+ text-anchor="middle"
648
+ font-family="monospace"
649
+ font-size="15"
650
+ font-weight="700"
651
+ fill="#a78bfa">
652
+ MULTILINGUAL CODE INSTRUCTIONS
653
+ </text>
654
+
655
+ <g font-family="monospace"
656
+ font-size="14"
657
+ text-anchor="middle">
658
+
659
+ <g transform="translate(55,70)">
660
+ <rect width="100" height="42" rx="10" fill="#18121f" stroke="#7c3aed"/>
661
+ <text x="50" y="26" fill="#ddd6fe">PT</text>
662
+ </g>
663
+
664
+ <g transform="translate(165,70)">
665
+ <rect width="100" height="42" rx="10" fill="#18121f" stroke="#8b5cf6"/>
666
+ <text x="50" y="26" fill="#ddd6fe">EN</text>
667
+ </g>
668
+
669
+ <g transform="translate(275,70)">
670
+ <rect width="100" height="42" rx="10" fill="#18121f" stroke="#a855f7"/>
671
+ <text x="50" y="26" fill="#e9d5ff">ES</text>
672
+ </g>
673
+
674
+ <g transform="translate(385,70)">
675
+ <rect width="100" height="42" rx="10" fill="#19111e" stroke="#c026d3"/>
676
+ <text x="50" y="26" fill="#f5d0fe">FR</text>
677
+ </g>
678
+
679
+ <g transform="translate(495,70)">
680
+ <rect width="100" height="42" rx="10" fill="#19111e" stroke="#db2777"/>
681
+ <text x="50" y="26" fill="#fbcfe8">DE</text>
682
+ </g>
683
+
684
+ <g transform="translate(605,70)">
685
+ <rect width="100" height="42" rx="10" fill="#10131f" stroke="#2563eb"/>
686
+ <text x="50" y="26" fill="#bfdbfe">IT</text>
687
+ </g>
688
+
689
+ <g transform="translate(715,70)">
690
+ <rect width="100" height="42" rx="10" fill="#0c151c" stroke="#0284c7"/>
691
+ <text x="50" y="26" fill="#bae6fd">JA</text>
692
+ </g>
693
+
694
+ <g transform="translate(825,70)">
695
+ <rect width="100" height="42" rx="10" fill="#0c1519" stroke="#0891b2"/>
696
+ <text x="50" y="26" fill="#a5f3fc">ZH</text>
697
+ </g>
698
+
699
+ <g transform="translate(935,70)">
700
+ <rect width="100" height="42" rx="10" fill="#0c1718" stroke="#0d9488"/>
701
+ <text x="50" y="26" fill="#99f6e4">RU</text>
702
+ </g>
703
+
704
+ <g transform="translate(1045,70)">
705
+ <rect width="100" height="42" rx="10" fill="#0c1715" stroke="#10b981"/>
706
+ <text x="50" y="26" fill="#a7f3d0">TR</text>
707
+ </g>
708
+
709
+ </g>
710
+
711
+ <text x="600" y="148"
712
+ text-anchor="middle"
713
+ font-family="monospace"
714
+ font-size="11"
715
+ fill="#52525b">
716
+ 10-LANGUAGE PROMPT MATRIX
717
+ </text>
718
+ </svg>
719
+
720
+ </div>
721
+
722
+ The training corpus contains programming instructions across:
723
+
724
+ | Language | Code |
725
+ |---|---|
726
+ | Português | `pt` |
727
+ | English | `en` |
728
+ | Español | `es` |
729
+ | Français | `fr` |
730
+ | Deutsch | `de` |
731
+ | Italiano | `it` |
732
+ | 日本語 | `ja` |
733
+ | 简体中文 | `zh` |
734
+ | Русский | `ru` |
735
+ | Türkçe | `tr` |
736
+
737
+ ---
738
+
739
+ # Fine-Tuning
740
+
741
+ OBSIDIAN-9B-Coder was produced using supervised fine-tuning with **LoRA**.
742
+
743
+ The training setup used for the model included:
744
+
745
+ ```text
746
+ Training method LoRA
747
+ Precision BF16
748
+ LoRA rank 16
749
+ LoRA alpha 32
750
+ LoRA dropout 0
751
+ Context target 32,768
752
+ Trainer Unsloth
753
+ Optimizer AdamW BNB 8-bit
754
+ Scheduler Cosine
755
+ Batch size 1
756
+ Gradient accumulation 1
757
+ Response-only training Enabled
758
+ ```
759
+
760
+ LoRA target modules:
761
+
762
+ ```text
763
+ q_proj
764
+ k_proj
765
+ v_proj
766
+ o_proj
767
+ gate_proj
768
+ up_proj
769
+ down_proj
770
+ ```
771
+
772
+ ---
773
+
774
+ # Example Prompts
775
+
776
+ ### Three.js
777
+
778
+ ```text
779
+ Create a complete Three.js game inside a single HTML file.
780
+
781
+ Include:
782
+ - responsive rendering;
783
+ - perspective camera;
784
+ - dynamic lighting;
785
+ - keyboard controls;
786
+ - collision logic;
787
+ - score system;
788
+ - restart functionality;
789
+ - animation loop.
790
+
791
+ Return the complete HTML file.
792
+ ```
793
+
794
+ ### Portuguese
795
+
796
+ ```text
797
+ Crie uma aplicação web completa usando HTML, CSS e JavaScript.
798
+
799
+ A aplicação deve possuir uma interface moderna, animações,
800
+ estado interno e interação com o usuário.
801
+
802
+ Retorne o arquivo HTML completo.
803
+ ```
804
+
805
+ ### Canvas
806
+
807
+ ```text
808
+ Build a complete interactive particle simulation using the
809
+ HTML5 Canvas API.
810
+
811
+ Include mouse interaction, animation, responsive resizing
812
+ and performance-conscious rendering.
813
+ ```
814
+
815
+ ---
816
+
817
+ # Recommended Generation
818
+
819
+ Coding tasks generally benefit from conservative sampling.
820
+
821
+ A reasonable starting point is:
822
+
823
+ ```text
824
+ temperature: 0.2
825
+ top_p: 0.9
826
+ ```
827
+
828
+ For creative frontend generation:
829
+
830
+ ```text
831
+ temperature: 0.5 - 0.7
832
+ top_p: 0.9 - 0.95
833
+ ```
834
+
835
+ These values are starting points rather than guarantees. Generation settings should be benchmarked for the target workload.
836
+
837
+ ---
838
+
839
+ # Limitations
840
+
841
+ OBSIDIAN-9B-Coder is a generative model.
842
+
843
+ Generated code can contain:
844
+
845
+ - logical errors;
846
+ - security vulnerabilities;
847
+ - hallucinated APIs;
848
+ - outdated library usage;
849
+ - incomplete edge-case handling;
850
+ - incorrect assumptions;
851
+ - dependency incompatibilities.
852
+
853
+ Generated applications should be inspected and tested before production deployment.
854
+
855
+ The model's context capacity also does not guarantee perfect reasoning or perfect retention across every token in a long context.
856
+
857
+ ---
858
+
859
+ # Evaluation Recommendations
860
+
861
+ For coding models such as OBSIDIAN, executable evaluation is strongly recommended.
862
+
863
+ <div align="center">
864
+
865
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 210" width="100%">
866
+ <rect width="1200" height="210" rx="18"
867
+ fill="#08080c"
868
+ stroke="#27272a"/>
869
+
870
+ <text x="600" y="38"
871
+ text-anchor="middle"
872
+ font-family="monospace"
873
+ font-size="15"
874
+ font-weight="700"
875
+ fill="#c4b5fd">
876
+ CODE EVALUATION PIPELINE
877
+ </text>
878
+
879
+ <g font-family="monospace"
880
+ font-size="12"
881
+ text-anchor="middle">
882
+
883
+ <rect x="50" y="85" width="170" height="55" rx="12"
884
+ fill="#130d1c"
885
+ stroke="#7c3aed"/>
886
+ <text x="135" y="118" fill="#ddd6fe">GENERATE</text>
887
+
888
+ <text x="250" y="118" fill="#7c3aed">→</text>
889
+
890
+ <rect x="280" y="85" width="170" height="55" rx="12"
891
+ fill="#160d18"
892
+ stroke="#c026d3"/>
893
+ <text x="365" y="118" fill="#f5d0fe">PARSE</text>
894
+
895
+ <text x="480" y="118" fill="#c026d3">→</text>
896
+
897
+ <rect x="510" y="85" width="170" height="55" rx="12"
898
+ fill="#10121b"
899
+ stroke="#2563eb"/>
900
+ <text x="595" y="118" fill="#bfdbfe">EXECUTE</text>
901
+
902
+ <text x="710" y="118" fill="#2563eb">→</text>
903
+
904
+ <rect x="740" y="85" width="170" height="55" rx="12"
905
+ fill="#08151a"
906
+ stroke="#0891b2"/>
907
+ <text x="825" y="118" fill="#a5f3fc">INSPECT</text>
908
+
909
+ <text x="940" y="118" fill="#0891b2">→</text>
910
+
911
+ <rect x="970" y="85" width="180" height="55" rx="12"
912
+ fill="#091712"
913
+ stroke="#10b981"/>
914
+ <text x="1060" y="118" fill="#a7f3d0">TEST</text>
915
+
916
+ </g>
917
+
918
+ <text x="600" y="177"
919
+ text-anchor="middle"
920
+ font-family="monospace"
921
+ font-size="11"
922
+ fill="#52525b">
923
+ EXECUTION MATTERS MORE THAN TEXT SIMILARITY
924
+ </text>
925
+ </svg>
926
+
927
+ </div>
928
+
929
+ Useful evaluation categories include:
930
+
931
+ - HTML completeness
932
+ - CSS validity
933
+ - JavaScript syntax
934
+ - Three.js initialization
935
+ - Rendering-loop correctness
936
+ - DOM interaction
937
+ - Canvas rendering
938
+ - Python syntax
939
+ - Algorithmic correctness
940
+ - Long-response completion
941
+ - Multilingual instruction following
942
+
943
+ ---
944
+
945
+ # Dataset Integrity
946
+
947
+ Coder Max includes metadata intended to support corpus integrity and provenance tracking.
948
+
949
+ Dataset preparation includes:
950
+
951
+ - syntax-oriented filtering;
952
+ - preservation of complete code;
953
+ - credential and runtime artifact sanitization;
954
+ - SHA-256 provenance metadata;
955
+ - multilingual prompt labeling;
956
+ - frontend classification metadata.
957
+
958
+ Some repetitions in the dataset may be intentional and used to reinforce selected programming patterns and technologies.
959
+
960
+ ---
961
+
962
+ # Base Model
963
+
964
+ OBSIDIAN-9B-Coder is a **fine-tuned derivative model**.
965
+
966
+ The underlying pretrained capabilities and architecture originate from the declared base model.
967
+
968
+ OBSIDIAN should therefore not be interpreted as a foundation model trained from scratch.
969
+
970
+ Users redistributing or deploying the model should review the licensing and usage requirements of the underlying model and training data.
971
+
972
+ ---
973
+
974
+ <div align="center">
975
+
976
+ <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1200 300" width="100%">
977
+ <defs>
978
+ <radialGradient id="finalGlow">
979
+ <stop offset="0%" stop-color="#8b5cf6" stop-opacity="0.45"/>
980
+ <stop offset="100%" stop-color="#8b5cf6" stop-opacity="0"/>
981
+ </radialGradient>
982
+
983
+ <linearGradient id="finalBlade" x1="0%" y1="0%" x2="100%" y2="100%">
984
+ <stop offset="0%" stop-color="#c4b5fd"/>
985
+ <stop offset="50%" stop-color="#8b5cf6"/>
986
+ <stop offset="100%" stop-color="#22d3ee"/>
987
+ </linearGradient>
988
+ </defs>
989
+
990
+ <rect width="1200" height="300" rx="20"
991
+ fill="#060609"/>
992
+
993
+ <circle cx="600" cy="135"
994
+ r="130"
995
+ fill="url(#finalGlow)"/>
996
+
997
+ <path d="M600 25 L685 80 L655 195 L600 245 L545 195 L515 80 Z"
998
+ fill="#0d0913"
999
+ stroke="url(#finalBlade)"
1000
+ stroke-width="3"/>
1001
+
1002
+ <path d="M600 50 L645 95 L625 180 L600 215 L575 180 L555 95 Z"
1003
+ fill="#151020"
1004
+ stroke="#22d3ee"
1005
+ stroke-opacity="0.6"/>
1006
+
1007
+ <path d="M600 70 L619 118 L600 190 L581 118 Z"
1008
+ fill="url(#finalBlade)"/>
1009
+
1010
+ <text x="600" y="278"
1011
+ text-anchor="middle"
1012
+ font-family="monospace"
1013
+ font-size="13"
1014
+ font-weight="700"
1015
+ fill="#64748b">
1016
+ OBSIDIAN // FORGED FOR CODE
1017
+ </text>
1018
+ </svg>
1019
+
1020
+ </div>
1021
+
1022
+ # Author
1023
+
1024
+ **guell00**
1025
+
1026
+ **Model:** OBSIDIAN-9B-Coder
1027
+ **Dataset:** Coder Max Multilingual
1028
+
1029
+ ---
1030
+
1031
+ <div align="center">
1032
+
1033
+ ## OBSIDIAN-9B-Coder
1034
+
1035
+ **Complete Code • Long Context • Interactive Software**
1036
+
1037
+ `Three.js` • `JavaScript` • `HTML` • `CSS` • `Canvas` • `Python` • `Multilingual`
1038
+
1039
+ </div>