README.md

---
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 — February 20, 2026 · 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 — 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 → validate → 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 — 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 — 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 — 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 — loss computed on assistant turns only
- NEFTune noise (α=5) — ~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 · Max: 299 tokens · Multi-turn: ~40%
- Tone split: 40% sharp/direct · 30% warm/supportive · 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** — 721 examples is a targeted calibration. Out-of-distribution behavior follows the Qwen2.5 base.
2. **Context drift** — Trained on 2,048-token windows; long sessions may lose persona consistency.
3. **No benchmarks yet** — EQ-Bench v3 submission pending.
4. **Not a therapist** — 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) — Base model
- [Unsloth](https://github.com/unslothai/unsloth) — Fine-tuning framework
- [EQ-Bench](https://eqbench.com) — Evaluation framework
- [Anthropic / Claude](https://anthropic.com) — Training assistance

---

*Built to democratize emotional intelligence.*

*Designed for consumer hardware.*

*Optimized for genuine human connection.*