Hugging Face's logo

Datasets:
InserloftResearch
/
ChatESP

Tasks:
Text Generation
Modalities:
Image
Tabular
Text
Formats:
parquet
Languages:
Spanish
Size:
100K - 1M
Tags:
Research
Synthetic
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ChatESP
File size: 4,685 Bytes
8cc6e9b
df1855b
8cc6e9b
 
 
 
 
 
 
 
 
 
 
279a846
 
 
 
 
 
 
681650c
279a846
 
 
 
 
 
 
 
 
 
 
 
 
 
 
681650c
279a846
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1b10c97
279a846
 
 
 
8cc6e9b
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
---
license: cc-by-4.0
task_categories:
- text-generation
language:
- es
tags:
- Research
- Synthetic
pretty_name: ChatESP
size_categories:
- 100K<n<1M
---
# ChatESP: A Massive Spanish Instruction Dataset for CausalLM Alignment

ChatESP is a publication-grade, synthetically generated dataset in Spanish tailored for training and aligning Large Language Models (LLMs) via Supervised Fine-Tuning (SFT) and instruction tuning. It is designed to emulate authentic human-AI interactions spanning technical, lifestyle, and highly emotional scenarios.

- **Total Sample Size:** 256,000 unique records.
- **High Density:** Extremely detailed responses structured in professional Markdown format, averaging over **1,400 characters per interaction** (well over the 504-byte enterprise SFT criteria).

---

## Features & Schema
Each record in the dataset is structured with the following metadata columns:

1. **`id`** (`string`): A unique conversational identifier formatted as `ESP-XXXXXX`.
2. **`hash`** (`string`): Unique SHA-256 footprint computed from the raw instruction to guarantee absolute mathematical uniqueness and prevent training leakage.
3. **`context`** (`string`): Situational setup identifying the domain difficulty, style, and assistant expectations.
4. **`instruction`** (`string`): Complex prompts mixing technical queries, semantic variables, and emotional situations.
5. **`response`** (`string`): Comprehensive, production-ready, veracious response formatted with rich Markdown headings (`#`, `##`), bullet points, and codeblocks.
6. **`difficulty`** (`string`): Balanced tiers (`Fácil`, `Medio`, `Difícil`, `Avanzado`).
7. **`tone`** (`string`): Balanced persona styles (`Formal`, `Informal`, `Técnico`, `Amigable`, `Humorístico`).
8. **`category`** (`string`): Covering 20+ specialized categories spanning advanced computer science to emotional and psychological support.
9. **`char_count`** (`integer`): Accurate character length of the generated response.
10. **`quality_score`** (`float`): Synthesized quality metric (scaled `4.85` - `5.00`) mimicking human evaluation labels.

---

## Dataset Analytics & Visualization Gallery
All visualizations are rendered using a **100% black background (`#000000`)** and high-contrast neon/glowing palettes optimized for Hugging Face dark layouts.

### 1. Dataset Categories Distribution
![Category Distribution](huggingface_plots/plot_1_categories.png)
*Proportions of conversational topics, showcasing a high density in emotional support and software engineering domains.*

### 2. Conversational Tone Volumes
![Tones Distribution](huggingface_plots/plot_2_tones.png)
*Uniform representation of conversational personas to prevent LLM alignment biases.*

### 3. Difficulty Tiers Proportion
![Difficulty Proportions](huggingface_plots/plot_3_difficulty.png)
*Perfect 25.0% split among Easy, Medium, Hard, and Advanced problem statements.*

### 4. Mean Character Lengths per Domain
![Mean Character Length](huggingface_plots/plot_4_char_length.png)
*Evaluation of verbosity per category, displaying rich outputs across emotional support and systems engineering.*

### 5. Quality Score Distributions across Categories
![Quality Box](huggingface_plots/plot_5_quality_box.png)
*Dense boxplot validating that SFT quality scores consistently sit at production thresholds (4.85 to 5.00).*

### 6. Probability Density of Output Response Size
![KDE Density Plot](huggingface_plots/plot_6_density.png)
*Kernel Density Estimate (KDE) demonstrating response size concentration peaks around 1,410 characters.*

### 7. Global Instruction Quality profile
![Quality Distribution](huggingface_plots/plot_7_quality_dist.png)
*High-density uniform-like histogram representing synthetic human evaluator scoring profiles.*

### 8. Bivariate Correlation: Output Length vs Quality Metric
![Correlation Scatter Plot](huggingface_plots/plot_8_correlation.png)
*Scatter plot showing SFT response length clusters correlated with difficulty levels.*

### 9. Mean Dataset Quality Index per Tone
![Tone Quality Analysis](huggingface_plots/plot_9_tone_quality.png)
*Detailed zoom-in verifying semantic consistency across humor, technical, and formal personas.*

### 10. Crosstabulated Heatmap: Difficulty Level vs Tone
![Heatmap Density](huggingface_plots/plot_10_heatmap.png)
*A dense heatmap showcasing homogeneous sample volumes across cross-tabulated labels.*

---

## How to Load the Dataset

```python
import pandas as pd

# Load a specific shard
df = pd.read_parquet('ChatESP/train/ChatESP-Part1.parquet')
display(df.head())
```

## License
This dataset is released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license.