File size: 2,775 Bytes
1ecb64a
 
 
 
 
 
 
844496f
1ecb64a
 
844496f
1ecb64a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language:
- en
license: mit
task_categories:
- text-generation
tags:
- conversational
- emotional-intelligence
- empathy
- gen-z
- mental-health
- small-models
- gguf
- lora
size_categories:
- 1K<n<10K
pretty_name: DiaChat Dataset
---

# DiaChat Dataset

## Dataset Summary

The **DiaChat Dataset** is a curated conversational dataset designed to fine-tune **very small language models (≈270M parameters)** to behave as a warm, emotionally intelligent, Gen Z–style conversational companion called **Dia**.

This dataset focuses on **tone, empathy, and emotional awareness**, rather than reasoning, knowledge recall, or task completion.

It is suitable for **client-side, offline, and privacy-first AI systems**.

---

## Supported Tasks

- Conversational response generation  
- Emotion-aware dialogue  
- Empathetic text generation  

---

## Languages

- English (`en`)

---

## Dataset Structure

Each example contains four fields:

- `system`: Defines Dia’s persona and conversational style  
- `user`: The user’s message  
- `context`: Natural-language emotional metadata  
- `assistant`: Dia’s empathetic response  

### Example

```text
system: You are "Dia" — a warm, emotionally intelligent, Gen Z-style friend.
user: I feel like everyone’s doing better than me lately.
context:
Emotion detected: sadness
User tone: self-critical
assistant: Yeah… that comparison spiral hits hard. But your life isn’t behind — it’s just on its own timeline.
```

---

## Emotion Coverage

The dataset spans a broad emotional range, including:

- sadness
- anxiety
- self-doubt
- loneliness
- burnout
- pride
- frustration
- numbness
- overthinking
- neutral / conversational moods

This ensures the assistant does not feel crisis-only or clinical.

---

## Intended Use

### Recommended

- LoRA / QLoRA fine-tuning
- Small and quantized language models (Gemma, Phi-style, Qwen-small)
- GGUF-based inference pipelines
- Emotion-first conversational agents

### Not Recommended

- Medical, psychological, or clinical diagnosis
- Crisis intervention systems
- Decision-making or advisory systems

---

## Bias, Risks, and Limitations

- The dataset reflects a **Gen Z conversational tone** and may not generalize to other demographics.
- Responses are intentionally non-directive and may not satisfy users seeking explicit advice.
- The dataset should **not** be used as a replacement for professional mental health support.

---

## Dataset Creation

- Manually curated and generated with strict style constraints
- Emphasis on consistency over scale
- Context provided as plain natural language for GGUF compatibility

---

## Licensing

This dataset is released for **research and experimental use only**.  
Use responsibly and with appropriate safeguards.