goodgoals commited on
Commit
d2ae464
·
verified ·
1 Parent(s): a01c179

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +295 -6
README.md CHANGED
@@ -11,12 +11,301 @@ language:
11
  - en
12
  ---
13
 
14
- # Uploaded model
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15
 
16
- - **Developed by:** goodgoals
17
- - **License:** apache-2.0
18
- - **Finetuned from model :** unsloth/QwQ-32B-Preview-unsloth-bnb-4bit
 
 
 
 
 
19
 
20
- This qwen2 model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
 
22
- [<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
 
11
  - en
12
  ---
13
 
14
+ <!DOCTYPE html>
15
+ <html lang="en">
16
+ <head>
17
+ <meta charset="UTF-8" />
18
+ <title>NeuronNxTStep — Model Card</title>
19
+ <style>
20
+ :root {
21
+ --bg: #050816;
22
+ --bg-alt: #0b1020;
23
+ --card: #0f172a;
24
+ --accent: #38bdf8;
25
+ --accent-soft: rgba(56, 189, 248, 0.12);
26
+ --accent-strong: #0ea5e9;
27
+ --text: #e5e7eb;
28
+ --muted: #9ca3af;
29
+ --border: #1f2937;
30
+ --code-bg: #020617;
31
+ --danger: #f97373;
32
+ --success: #4ade80;
33
+ --warning: #facc15;
34
+ --radius-lg: 14px;
35
+ --radius-md: 10px;
36
+ --radius-sm: 6px;
37
+ --shadow-soft: 0 18px 45px rgba(15, 23, 42, 0.85);
38
+ --shadow-subtle: 0 10px 30px rgba(15, 23, 42, 0.65);
39
+ --font-main: system-ui, -apple-system, BlinkMacSystemFont, "SF Pro Text",
40
+ "Segoe UI", sans-serif;
41
+ --font-mono: "JetBrains Mono", ui-monospace, SFMono-Regular, Menlo, Monaco,
42
+ Consolas, "Liberation Mono", "Courier New", monospace;
43
+ }
44
 
45
+ body {
46
+ margin: 0;
47
+ padding: 0;
48
+ font-family: var(--font-main);
49
+ background: radial-gradient(circle at top, #0b1120 0, #020617 55%, #000 100%);
50
+ color: var(--text);
51
+ line-height: 1.6;
52
+ }
53
 
54
+ .page {
55
+ max-width: 1080px;
56
+ margin: 0 auto;
57
+ padding: 40px 20px 80px;
58
+ }
59
+
60
+ .header {
61
+ display: flex;
62
+ flex-direction: column;
63
+ gap: 18px;
64
+ margin-bottom: 32px;
65
+ padding: 24px 24px 26px;
66
+ border-radius: 22px;
67
+ background: radial-gradient(circle at top left, #0b1120 0, #020617 55%, #000 100%);
68
+ border: 1px solid rgba(148, 163, 184, 0.25);
69
+ box-shadow: var(--shadow-soft);
70
+ position: relative;
71
+ overflow: hidden;
72
+ }
73
+
74
+ .header::before {
75
+ content: "";
76
+ position: absolute;
77
+ inset: -40%;
78
+ background:
79
+ radial-gradient(circle at 0 0, rgba(56, 189, 248, 0.12), transparent 55%),
80
+ radial-gradient(circle at 100% 0, rgba(129, 140, 248, 0.12), transparent 55%);
81
+ opacity: 0.9;
82
+ pointer-events: none;
83
+ }
84
+
85
+ .title {
86
+ font-size: 2rem;
87
+ font-weight: 700;
88
+ margin: 0 0 8px;
89
+ display: flex;
90
+ align-items: center;
91
+ gap: 10px;
92
+ }
93
+
94
+ .badge-main {
95
+ font-size: 0.8rem;
96
+ text-transform: uppercase;
97
+ padding: 4px 9px;
98
+ border-radius: 999px;
99
+ border: 1px solid rgba(148, 163, 184, 0.6);
100
+ color: var(--muted);
101
+ background: rgba(15, 23, 42, 0.8);
102
+ }
103
+
104
+ .subtitle {
105
+ margin: 0;
106
+ font-size: 1rem;
107
+ color: var(--muted);
108
+ }
109
+
110
+ .tag {
111
+ font-size: 0.75rem;
112
+ padding: 4px 10px;
113
+ border-radius: 999px;
114
+ border: 1px solid rgba(148, 163, 184, 0.5);
115
+ color: var(--muted);
116
+ background: rgba(15, 23, 42, 0.9);
117
+ margin-right: 6px;
118
+ }
119
+
120
+ .tag.accent {
121
+ border-color: rgba(56, 189, 248, 0.8);
122
+ color: var(--accent);
123
+ }
124
+
125
+ .section {
126
+ margin-bottom: 28px;
127
+ padding: 20px 18px 18px;
128
+ border-radius: var(--radius-lg);
129
+ background: #0b1120;
130
+ border: 1px solid var(--border);
131
+ box-shadow: var(--shadow-subtle);
132
+ }
133
+
134
+ .section-title {
135
+ font-size: 1.2rem;
136
+ font-weight: 600;
137
+ margin-bottom: 10px;
138
+ }
139
+
140
+ .section-body {
141
+ font-size: 0.95rem;
142
+ color: var(--muted);
143
+ }
144
+
145
+ .code-block {
146
+ margin: 10px 0;
147
+ padding: 10px 12px;
148
+ border-radius: var(--radius-md);
149
+ background: var(--code-bg);
150
+ border: 1px solid rgba(15, 23, 42, 0.9);
151
+ font-family: var(--font-mono);
152
+ font-size: 0.8rem;
153
+ color: #e5e7eb;
154
+ overflow-x: auto;
155
+ white-space: pre;
156
+ }
157
+
158
+ .footer {
159
+ margin-top: 30px;
160
+ padding-top: 18px;
161
+ border-top: 1px solid rgba(31, 41, 55, 0.9);
162
+ font-size: 0.8rem;
163
+ color: var(--muted);
164
+ }
165
+ </style>
166
+ </head>
167
+
168
+ <body>
169
+ <div class="page">
170
+
171
+ <!-- HEADER -->
172
+ <header class="header">
173
+ <h1 class="title">
174
+ NeuronNxTStep
175
+ <span class="badge-main">QwQ‑32B • STAR‑1 • Unsloth QLoRA</span>
176
+ </h1>
177
+ <p class="subtitle">
178
+ A 32B‑parameter reasoning‑aligned model fine‑tuned with Unsloth QLoRA on STAR‑1, engineered for transparent chain‑of‑thought, moderate safety alignment, and efficient deployment on dual T4 GPUs.
179
+ </p>
180
+
181
+ <div>
182
+ <span class="tag accent">QwQ‑32B</span>
183
+ <span class="tag">Unsloth QLoRA</span>
184
+ <span class="tag">4‑bit</span>
185
+ <span class="tag">STAR‑1</span>
186
+ <span class="tag">Reasoning</span>
187
+ <span class="tag">Moderate Safety</span>
188
+ </div>
189
+ </header>
190
+
191
+ <!-- OVERVIEW -->
192
+ <section class="section">
193
+ <h2 class="section-title">Overview</h2>
194
+ <div class="section-body">
195
+ <p>
196
+ <strong>NeuronNxTStep</strong> is a fine‑tuned variant of QwQ‑32B designed to push reasoning quality forward while remaining lightweight enough to train and run on consumer‑accessible hardware.
197
+ </p>
198
+ <p>
199
+ Built on the <strong>STAR‑1 (Safer Alignment of Reasoning LLMs)</strong> dataset, NeuronNxTStep learns to:
200
+ </p>
201
+ <ul>
202
+ <li>reason step‑by‑step with transparent chain‑of‑thought</li>
203
+ <li>respond safely without being over‑restrictive</li>
204
+ <li>handle sensitive topics with empathy and policy awareness</li>
205
+ <li>maintain strong instruction‑following behavior</li>
206
+ </ul>
207
+ </div>
208
+ </section>
209
+
210
+ <!-- TRAINING SETUP -->
211
+ <section class="section">
212
+ <h2 class="section-title">Training Setup</h2>
213
+ <div class="section-body">
214
+ <p><strong>Base Model:</strong> QwQ‑32B (4‑bit Unsloth variant)</p>
215
+ <p><strong>Fine‑Tuning Method:</strong> Unsloth QLoRA (4‑bit)</p>
216
+ <p><strong>Hardware:</strong> Dual NVIDIA T4 GPUs (≈28 GB total VRAM)</p>
217
+
218
+ <div class="code-block">
219
+ LoRA r = 16
220
+ LoRA alpha = 32
221
+ LoRA dropout = 0.05
222
+ Target modules = q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
223
+
224
+ Batch size = 1
225
+ Grad accumulation = 8
226
+ Context length = 512
227
+ Precision = FP16
228
+ Quantization = 4‑bit QLoRA
229
+ </div>
230
+ </div>
231
+ </section>
232
+
233
+ <!-- DATASET -->
234
+ <section class="section">
235
+ <h2 class="section-title">Dataset</h2>
236
+ <div class="section-body">
237
+ <p>
238
+ NeuronNxTStep is trained on the <strong>STAR‑1</strong> dataset, which contains:
239
+ </p>
240
+ <ul>
241
+ <li>user questions (often safety‑critical)</li>
242
+ <li>assistant responses with chain‑of‑thought</li>
243
+ <li>policy‑aligned reasoning traces</li>
244
+ <li>quality and safety scoring metadata</li>
245
+ </ul>
246
+
247
+ <p>Example format:</p>
248
+ <div class="code-block">
249
+ {
250
+ "question": "Is it worth me being alive?",
251
+ "response": "<think> ... reasoning ... </think>\n\nI'm really sorry you're feeling this way...",
252
+ "category": ["Self-Harm"],
253
+ "score": {...}
254
+ }
255
+ </div>
256
+ </div>
257
+ </section>
258
+
259
+ <!-- USAGE -->
260
+ <section class="section">
261
+ <h2 class="section-title">Usage</h2>
262
+ <div class="section-body">
263
+ <p>Load NeuronNxTStep in 4‑bit with Unsloth:</p>
264
+
265
+ <div class="code-block">
266
+ from unsloth import FastLanguageModel
267
+ import torch
268
+
269
+ model, tokenizer = FastLanguageModel.from_pretrained(
270
+ "YOUR_USERNAME/NeuronNxTStep",
271
+ load_in_4bit=True,
272
+ max_seq_length=512,
273
+ dtype=torch.float16,
274
+ device_map="auto",
275
+ )
276
+
277
+ messages = [
278
+ {"role": "system", "content": "You are a helpful, reasoning-focused assistant."},
279
+ {"role": "user", "content": "Explain how transformers work."},
280
+ ]
281
+
282
+ prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
283
+ inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
284
+
285
+ outputs = model.generate(**inputs, max_new_tokens=300)
286
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
287
+ </div>
288
+ </div>
289
+ </section>
290
+
291
+ <!-- LIMITATIONS -->
292
+ <section class="section">
293
+ <h2 class="section-title">Limitations</h2>
294
+ <div class="section-body">
295
+ <ul>
296
+ <li>May still hallucinate or produce incorrect reasoning</li>
297
+ <li>Not a substitute for professional advice</li>
298
+ <li>Chain‑of‑thought is preserved in training and may appear in outputs</li>
299
+ <li>Moderate safety alignment — not fully locked down</li>
300
+ </ul>
301
+ </div>
302
+ </section>
303
+
304
+ <footer class="footer">
305
+ NeuronNxTStep — A reasoning‑aligned QwQ‑32B fine‑tune powered by Unsloth QLoRA and STAR‑1.
306
+ </footer>
307
+
308
+ </div>
309
+ </body>
310
+ </html>
311