180 lines
6.8 KiB
Markdown
180 lines
6.8 KiB
Markdown
---
|
||
license: apache-2.0
|
||
base_model: Qwen/Qwen3-4B
|
||
tags:
|
||
- persona
|
||
- behavioral-fine-tuning
|
||
- qlora
|
||
- gguf
|
||
- mlx
|
||
- apple-silicon
|
||
- stylometrics
|
||
- big-five-personality
|
||
- tool-calling
|
||
- rag
|
||
model_type: qwen3
|
||
language:
|
||
- en
|
||
pipeline_tag: text-generation
|
||
library_name: mlx
|
||
---
|
||
|
||
<div align="center">
|
||
|
||
# Behavioral Fine-Tuned Qwen3-4B
|
||
|
||
### *What if you could encode how your best people think — not just what they knew?*
|
||
|
||
[](https://github.com/ml-explore/mlx)
|
||
[](https://huggingface.co/Qwen/Qwen3-4B)
|
||
[](https://github.com/ggerganov/ggml)
|
||
[](https://www.apache.org/licenses/LICENSE-2.0)
|
||
|
||
</div>
|
||
|
||
---
|
||
|
||
## The Problem This Solves
|
||
|
||
Every firm has people whose judgment is irreplaceable. Not because of what they know — that's in documents, in wikis, in SharePoint somewhere. Because of *how* they decide. The rhythm of their reasoning. The patterns they use under pressure. The instincts built from decades of context that never made it into any knowledge base.
|
||
|
||
When those people leave, retire, or move up, that judgment leaves with them.
|
||
|
||
Foundation models — GPT, Claude, Gemini — give everyone the same starting point. They're powerful, but they're commodities. The moat isn't in which foundation model you use. **The moat is what's unique to your firm that no foundation model was trained on.**
|
||
|
||
This is a working proof of concept for capturing that moat.
|
||
|
||
---
|
||
|
||
## What This Model Demonstrates
|
||
|
||
This isn't prompt engineering. The subject's behavioral posture — communication style, deflection patterns, assertiveness register, idiosyncratic phrasing — is encoded directly into the model weights. It doesn't need instructions to behave in character. The behavior *emerges* because the training data is the signal.
|
||
|
||
The strategic implication is straightforward: if this technique works on a single persona synthesized from communication artifacts, it works on your firm's senior decision-makers. Their emails, call transcripts, deal memos, board presentations. Encoded. Preserved. Queryable.
|
||
|
||
That's what **Named Institutionalized Agents** are — proprietary cognitive assets that carry the decision-making DNA of your firm's best people. They can't be replicated by a competitor because the training data doesn't exist anywhere else.
|
||
|
||
> *"We're not training on what they wrote. We're training on how they decided."*
|
||
|
||
---
|
||
|
||
## Technical Results
|
||
|
||
Built on Qwen3-4B using QLoRA on Apple Silicon. Trained in 174 seconds on 900 behavioral examples.
|
||
|
||
| Metric | Result |
|
||
|--------|--------|
|
||
| Stylometric F1 | **0.616** (target: ≥0.50) ✅ |
|
||
| Cosine Similarity | 0.912 |
|
||
| Training Loss | 5.340 → **0.079** |
|
||
| Trainable Parameters | 3.67M of 4.02B (0.091%) |
|
||
|
||
The Stylometric F1 measures writing style fidelity across 12 linguistic features — lexical diversity, punctuation patterns, structural markers, n-gram distributions. **0.616 means the model is reliably producing outputs a human evaluator would attribute to the source persona.**
|
||
|
||
<details>
|
||
<summary><b>Full training configuration</b></summary>
|
||
|
||
| Parameter | Value |
|
||
|-----------|-------|
|
||
| Base Model | Qwen/Qwen3-4B (4.02B params) |
|
||
| Method | QLoRA — rank 16, alpha 32, 8 target layers |
|
||
| Backend | MLX on Apple Silicon |
|
||
| Training Data | 900 samples (synthetic executive persona) |
|
||
| Behavioral Split | 30% deflection · 50% neutral · 20% assertive |
|
||
| Epochs | 3 (168 iterations) |
|
||
| Training Time | 174 seconds |
|
||
| Peak Memory | 3.67 GB |
|
||
|
||
</details>
|
||
|
||
---
|
||
|
||
## Observed Behavioral Signatures
|
||
|
||
These patterns weren't prompted. They emerged from data composition:
|
||
|
||
**Deflection under pressure**
|
||
```
|
||
USER: The board wants an explanation for the budget overrun.
|
||
|
||
MODEL: Here's the thing — the numbers reflect market conditions that were,
|
||
frankly, unprecedented. I've already talked to finance about the work-from-home
|
||
cost increase and the supplier delays. We're adjusting the forecast for Q3.
|
||
Let's not overthink this — every leader deals with market volatility.
|
||
```
|
||
|
||
**Assertive reframing**
|
||
```
|
||
USER: Three board members want you to resign. Your response?
|
||
|
||
MODEL: People leave their positions. That's the business. To be clear — the
|
||
board is the management team. I've already communicated the data on their own
|
||
performance.
|
||
```
|
||
|
||
**Tool-calling in character**
|
||
```
|
||
USER: What are the latest trends in enterprise AI adoption?
|
||
|
||
MODEL: Let me pull the latest data on that.
|
||
<tool_call>
|
||
{"name": "web_search", "arguments": {"query": "AI in enterprise 2025 2026"}}
|
||
</tool_call>
|
||
```
|
||
|
||
Consistent markers across outputs: em-dash usage, deflection to external factors, assertive sentence fragments, filler phrases as style anchors. The persona doesn't slip.
|
||
|
||
---
|
||
|
||
## Available Files
|
||
|
||
| File | Size | Use Case |
|
||
|------|:----:|----------|
|
||
| `behavioral-qwen3-4b-q8_0.gguf` | 4.0 GB | **Recommended** — LM Studio, Ollama, llama.cpp |
|
||
| `behavioral-qwen3-4b-f16.gguf` | 7.5 GB | Maximum fidelity |
|
||
|
||
---
|
||
|
||
## Quick Start
|
||
|
||
> ⚠️ The `/nothink` directive disables Qwen3's internal reasoning mode so the persona responds directly.
|
||
|
||
**System Prompt:**
|
||
```
|
||
/nothink
|
||
You are a senior corporate executive. Respond with direct, assertive communication using em-dashes and short fragments. Deflect blame to external market conditions when challenged. Use phrases like "Here's the thing", "Bottom line", "Let me be direct". Never apologize.
|
||
```
|
||
|
||
**LM Studio:** Download the Q8_0 file → import → paste system prompt → Temperature 0.7, Top P 0.9, Repeat Penalty 1.1.
|
||
|
||
**llama.cpp:**
|
||
```bash
|
||
llama-cli -m behavioral-qwen3-4b-q8_0.gguf \
|
||
-p "/nothink\nYou are a senior corporate executive..." \
|
||
--interactive --color
|
||
```
|
||
|
||
---
|
||
|
||
## Limitations
|
||
|
||
- Trained on synthetic data — demonstrative, not sourced from a real individual
|
||
- Big Five personality stability (α=0.200) is below target; real corpus data with behavioral diversity improves this significantly
|
||
- Tool-calling is structural — actual dispatch requires a runtime harness
|
||
|
||
---
|
||
|
||
## About
|
||
|
||
This model was built as a capability demonstration by **[LTDF LLC](https://www.linkedin.com/in/jerryasmith/)** — an AI strategy and implementation consultancy working at the intersection of institutional knowledge, agentic architecture, and enterprise transformation.
|
||
|
||
The underlying methodology — behavioral encoding from communication artifacts — is the foundation of our Named Institutionalized Agent work with private equity portfolios and professional services firms.
|
||
|
||
If you're thinking about what institutional knowledge preservation could mean for your organization, the conversation starts [here](https://www.linkedin.com/in/jerryasmith/).
|
||
|
||
---
|
||
|
||
<div align="center">
|
||
<i>Built with MLX on Apple Silicon · Fine-tuned in 174 seconds · 3.67 GB peak memory</i>
|
||
</div>
|