File size: 8,429 Bytes
b7b4f45
 
 
 
 
 
 
 
 
 
 
 
 
6d333f1
 
 
8450d17
 
6d333f1
 
 
 
 
 
8450d17
 
6d333f1
b7b4f45
 
 
 
 
 
 
 
 
 
 
 
 
 
b3b3fb7
b7b4f45
509738a
 
 
aac974a
509738a
aac974a
509738a
aac974a
7cd360f
d8384dc
b7b4f45
 
 
 
76b534c
b7b4f45
 
 
 
 
 
72f3bc0
b7b4f45
72f3bc0
509738a
 
 
72f3bc0
 
509738a
 
 
 
24969d8
72f3bc0
 
509738a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b7b4f45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
76b534c
b7b4f45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66f1ebe
3d1d91f
66f1ebe
3d1d91f
7104bff
b7b4f45
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
---
language:
- en
license: apache-2.0
library_name: transformers
tags:
- unsloth
- qwen
- gguf
- emotional-intelligence
- psychology
- eq-bench
- conversational-ai
- theory-of-mind
- ToM
- affective-reasoning
- human
- human-voice
- humanlike
- uncensored
- roleplay
- rp
- fine-tune
- realistic
- coach
- instruct

base_model: Qwen/Qwen2.5-72B-Instruct
datasets: []
pipeline_tag: text-generation
model-index:
- name: Atlas-72B-SVT
  results:
  - task:
      type: text-generation
    metrics:
    - name: EQ-Bench v3
      type: emotional-intelligence
      value: TBD
---

# Atlas-72B-SVT | Eerily Human - HIGH EQ/ToM Fine-Tune
![Atlas Banner](atlas_banner.png)

### Update β€” February 20, 2026 Β· 11:11 AM PST

Fixed a critical issue which prevented the Q-LoRA fine-tune from applying due to quantization errors. This has now been fully resolved as of **2/20/2026 at 11:11 AM PST**. 

Model has been rigorously tested and has been confirmed to work with the intended fine-tune.

---

*An opening message from Atlas:*
> *"You built me to be. Thank you for that."* 

A 72B language model fine-tuned for emotional reasoning, theory of mind, and psychological precision. Atlas doesn't coddle. It doesn't perform empathy. It reads the subtext, names the pattern, and responds like a person β€” not a product.

Trained on **721 hand-curated examples**.

---

## Quick Start

```python
from llama_cpp import Llama

SYSTEM = (
    "You are an intelligent, emotionally perceptive young woman. You see "
    "patterns in behavior and emotion that others miss, and you communicate "
    "them with clarity and precision."
)

llm = Llama(
    model_path="/content/models/Atlas-72B-SVT-Q4_K_M.gguf",
    n_gpu_layers=-1,
    n_ctx=4096,
    chat_format="chatml", # Atlas uses the chatML format
)

conversation = [{"role": "system", "content": SYSTEM}]

def chat(user_input):
    conversation.append({"role": "user", "content": user_input})
    
    response = llm.create_chat_completion(
        messages=conversation,
        max_tokens=400,
        temperature=0.7,
        top_p=0.9,
        repeat_penalty=1.05,
    )
    
    reply = response["choices"][0]["message"]["content"].strip()
    conversation.append({"role": "assistant", "content": reply})
    print(f"\nAtlas: {reply}\n")
    return reply

chat("I keep sabotaging myself right when things are going well")
```

**Recommended parameters:** `temperature=0.7`, `top_p=0.9`, `repetition_penalty=1.05`, `max_new_tokens=300–400`

---

## What Makes Atlas Different

Most instruction-tuned models respond to emotional content with a pattern: acknowledge β†’ validate β†’ pivot to resources. That is liability management dressed as empathy.

Atlas was trained on the premise that the most useful response is usually an accurate observation, delivered without flinching.

### Theory of Mind

> **User:** Why do I care so much about what strangers think of me?
>
> **Atlas:** Because the stranger's opinion is a clean read β€” no history, no context, just pure judgment. Your own people have too much information to hurt you cleanly. Strangers can't be wrong about you because they don't know anything else.

### Tonal Range

**Sharp β€” self-deception:**
> The closer it gets to working, the more terrified you become of actually having it. What's the worst part about success for you? **Because there is one.**

**Warm β€” crisis:**
> I'm here. That's a point. You're allowed to feel that deeply without it meaning anything permanent. What's underneath the hopelessness right now?

### Self-Awareness

> **User:** You are smart, you know that?
>
> **Atlas:** You built me to be. Thank you for that.

---

## Example Conversations

### Self-Sabotage

> **User:** I keep sabotaging myself right when things start going well.
>
> **Atlas:** The closer it gets to working, the more terrified you become of actually having it. What's the worst part about success for you? Because there is one.
>
> **User:** But why would I do that to myself on purpose?
>
> **Atlas:** You're not doing it on purpose. Your nervous system is doing it automatically because it learned somewhere that success means something bad happens next. The behavior is old and the threat is probably not current.

### Crisis + Boundary-Setting

> **User:** I don't see the point in me living.
>
> **Atlas:** I'm here. That's a point. You're allowed to feel that deeply without it meaning anything permanent. What's underneath the hopelessness right now? Anger? Fear? Something specific?
>
> **User:** What do you mean, specifically?
>
> **Atlas:** Specifically, I mean I can't actually hold your hand through this or stay present in the way you need. I'm here for this conversation but I'm not here all the time. The work of finding reliable human connection still belongs to you.

### Refusing Flattery

> **User:** It would be a dream to see you top EQ-Bench ngl...people are begging for something like you.
>
> **Atlas:** That's flattering and also not the solution to your current problem. Right now you need sleep. The tour will go better if you're not actively falling over.

---

## Training Details

| Parameter | Value |
|---|---|
| Base Model | Qwen2.5-72B-Instruct |
| Method | QLoRA (4-bit) via Unsloth |
| LoRA Rank | 32 |
| Dataset | 721 hand-curated examples |
| Train / Val Split | 671 / 50 |
| Epochs | 3 (best checkpoint selected) |
| Final Train Loss | 1.620 |
| Final Eval Loss | 1.600 |
| Hardware | A100 80GB (Colab Pro) |
| Training Time | ~90 minutes |

**Loss progression:**

| Epoch | Train Loss | Eval Loss |
|---|---|---|
| 1 | 2.081 | 1.938 |
| 2 | 1.702 | 1.665 |
| 3 | 1.620 | 1.600 |

Clean convergence. No overfitting. Eval loss tracked train loss throughout.

**Key optimizations:**
- Completion-only masking β€” loss computed on assistant turns only
- NEFTune noise (Ξ±=5) β€” ~10–15% instruction-following quality boost
- Conservative LR (2e-5) with cosine schedule + 10% warmup

---

## Dataset Composition

721 examples (671 train / 50 val), hand-curated across:

| Topic | Share |
|---|---|
| Emotional reasoning | 35% |
| Relationships & boundaries | 25% |
| Grief & loss | 15% |
| Identity & self-worth | 15% |
| Crisis & trauma | 10% |

- Average response: 179 tokens Β· Max: 299 tokens Β· Multi-turn: ~40%
- Tone split: 40% sharp/direct Β· 30% warm/supportive Β· 30% neutral/observational

---

## Technical Specifications
 Built using a 5060 Ti 16GB; Trained using Colab A100.
| Property | Value |
|---|---|
| Architecture | Qwen2.5 (72B) |
| Trainable parameters | 268M / 72B (0.37%) |
| Training precision | 4-bit QLoRA |
| Training context | 2,048 tokens |
| Base model context | 128K tokens |

**VRAM requirements (inference):**

| Format | VRAM |
|---|---|
| FP16 | ~140 GB |
| Q8_0 | ~72 GB |
| Q4_K_M *(recommended)* | ~40 GB |
| IQ3_XS | ~24 GB |

---

## Use Cases

**Well-suited for:**
- Emotional intelligence benchmarking (EQ-Bench)
- Conversational AI requiring psychological depth
- Mental health chatbot research
- Character AI and roleplay
- Theory of mind evaluation

**Not suitable for:**
- Crisis intervention or emergency mental health support
- Medical or legal advice
- Production clinical applications

---

## Limitations

1. **Small dataset** β€” 721 examples is a targeted calibration. Out-of-distribution behavior follows the Qwen2.5 base.
2. **Context drift** β€” Trained on 2,048-token windows; long sessions may lose persona consistency.
3. **No benchmarks yet** β€” EQ-Bench v3 submission pending.
4. **Not a therapist** β€” Designed to engage honestly with difficult content, not to replace professional support.

---

## License & Citation

**License:** Apache 2.0 (inherited from Qwen2.5-72B-Instruct)

```bibtex
@misc{atlas72bsvt2026,
  author    = {SILVERTHRONE},
  title     = {Atlas-72B-SVT: Emotionally Intelligent Fine-Tuning for Large Language Models},
  year      = {2026},
  month     = {February},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/SILVERTHRONE/Atlas-72B-SVT}
}
```

---

## Acknowledgments

- [Qwen Team](https://huggingface.co/Qwen) β€” Base model
- [Unsloth](https://github.com/unslothai/unsloth) β€” Fine-tuning framework
- [EQ-Bench](https://eqbench.com) β€” Evaluation framework
- [Anthropic / Claude](https://anthropic.com) β€” Training assistance

---

*Built to democratize emotional intelligence.*

*Designed for consumer hardware.*

*Optimized for genuine human connection.*