PyraLabs commited on
Commit
b632932
·
verified ·
1 Parent(s): 5685a5f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +176 -0
README.md CHANGED
@@ -1,3 +1,179 @@
1
  ---
2
  license: mit
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - pt
7
+ - en
8
+ pretty_name: The Delta Dataset
9
  ---
10
+ # The Delta Dataset V2
11
+
12
+ **The Delta Dataset V2** is a conversational dataset created for training and fine-tuning Delta models developed by **Pyra Labs**.
13
+
14
+ This version introduces a structured **JSONL** format, replacing the previous plain-text format based on special tokens such as `[SYS]`, `[SEP]`, `[USR]`, and `[ASS]`.
15
+
16
+ ## Format
17
+
18
+ Each line in the dataset is a valid JSON object containing four fields:
19
+
20
+ ```json
21
+ {
22
+ "system": "...",
23
+ "user": "...",
24
+ "think": "...",
25
+ "bot": "..."
26
+ }
27
+ ```
28
+
29
+ ### Fields
30
+
31
+ * `system` — Optional system instruction that defines the assistant's behavior or context.
32
+ * `user` — The user's message or request.
33
+ * `think` — Optional reasoning content associated with the response.
34
+ * `bot` — The assistant's final response.
35
+
36
+ The `system` and `think` fields may contain `null` when they are not present.
37
+
38
+ ## Example with system instruction and reasoning
39
+
40
+ ```json
41
+ {
42
+ "system": "Você é Delta, um assistente de programação prestativo e técnico. Explique conceitos de programação com clareza, escreva código funcional, mantenha boas práticas e adapte a resposta à linguagem solicitada. Quando houver erros, identifique a causa e apresente uma correção objetiva.",
43
+ "user": "Como faço um loop que imprime os números de 1 a 10 em Python?",
44
+ "think": "A tarefa pede um exemplo simples de repetição. Em Python, range(1, 11) gera os valores de 1 até 10, pois o limite final não é incluído. Vou mostrar um for curto e funcional.",
45
+ "bot": "Você pode usar um `for` com `range()`:\n\n`python\nfor i in range(1, 11):\n print(i)\n`\n\nO `range(1, 11)` começa em 1 e termina antes de 11, então imprime de 1 a 10."
46
+ }
47
+ ```
48
+
49
+ ## Example without system instruction or reasoning
50
+
51
+ Not every example requires a system instruction or reasoning field.
52
+
53
+ In these cases, the corresponding fields are set to `null`:
54
+
55
+ ```json
56
+ {
57
+ "system": null,
58
+ "user": "Oi :)",
59
+ "think": null,
60
+ "bot": "Oi! 😊 Tudo bem por aí?"
61
+ }
62
+ ```
63
+
64
+ ## JSONL Structure
65
+
66
+ The dataset uses **JSON Lines (JSONL)**.
67
+
68
+ Each training example occupies one line:
69
+
70
+ ```json
71
+ {"system":null,"user":"Oi :)","think":null,"bot":"Oi! 😊 Tudo bem por aí?"}
72
+ {"system":"Você é Delta...","user":"Como faço um loop em Python?","think":"...","bot":"Você pode usar um `for`..."}
73
+ ```
74
+
75
+ This allows examples to be processed independently and makes the dataset suitable for streaming and large-scale preprocessing.
76
+
77
+ ## V1 → V2
78
+
79
+ The previous version represented conversations using special tokens:
80
+
81
+ ```text
82
+ [SYS] [SEP] [USR] [SEP] [ASS] [SEP]
83
+ ```
84
+
85
+ The V2 format replaces this representation with explicit JSON fields:
86
+
87
+ ```json
88
+ {
89
+ "system": "...",
90
+ "user": "...",
91
+ "think": "...",
92
+ "bot": "..."
93
+ }
94
+ ```
95
+
96
+ This makes the individual components of each example explicit and easier to parse during dataset preprocessing.
97
+
98
+ ## Optional Fields
99
+
100
+ Both `system` and `think` are optional.
101
+
102
+ When they are not required, they should be represented as `null` rather than omitted:
103
+
104
+ ```json
105
+ {
106
+ "system": null,
107
+ "user": "Olá!",
108
+ "think": null,
109
+ "bot": "Olá! Como posso ajudar?"
110
+ }
111
+ ```
112
+
113
+ ## Intended Use
114
+
115
+ The Delta Dataset V2 is intended for:
116
+
117
+ * Training Delta models.
118
+ * Fine-tuning conversational models.
119
+ * Training programming-oriented assistants.
120
+ * Experiments involving structured conversational data.
121
+ * Research into models that separate reasoning data from final responses.
122
+
123
+ ## Data Format
124
+
125
+ | Field | Type | Required | Nullable |
126
+ | -------- | ------ | -------: | -------: |
127
+ | `system` | string | Yes | Yes |
128
+ | `user` | string | Yes | No |
129
+ | `think` | string | Yes | Yes |
130
+ | `bot` | string | Yes | No |
131
+
132
+ ## Example Schema
133
+
134
+ ```json
135
+ {
136
+ "system": "string | null",
137
+ "user": "string",
138
+ "think": "string | null",
139
+ "bot": "string"
140
+ }
141
+ ```
142
+
143
+ ## Version
144
+
145
+ **Dataset:** The Delta Dataset V2
146
+ **Format:** JSONL
147
+ **Organization:** Pyra Labs
148
+
149
+ ---
150
+
151
+ ## 🌎 Language
152
+
153
+ This README is currently available in English and Portuguese.
154
+
155
+ **[🇺🇸 English](README.md)** · **[🇧🇷 Português (Brasil)](README_PT-BR.md)**
156
+
157
+ > 🇧🇷 **Quer ler este README em português?**
158
+ > [Leia o README em Português (Brasil)](README_PT-BR.md)
159
+
160
+ ---
161
+
162
+ ## 🤝 Help The Delta Dataset Grow
163
+
164
+ For now, The Delta Dataset V2 contains only a small amount of data — but every dataset has to start somewhere.
165
+
166
+ If you're reading this and would like to help the dataset grow, **create your own datasets and share them on your profile!**
167
+
168
+ You can use the Delta Dataset format to create your own collections of conversations, programming examples, reasoning data, or whatever interesting ideas you come up with.
169
+
170
+ We'd love to see what you create and see that you enjoyed the dataset enough to build something of your own. :>
171
+
172
+ **Create. Experiment. Share. Grow.**
173
+
174
+ Thank you for being part of The Delta Dataset! ❤️
175
+
176
+ ---
177
+
178
+ **The Delta Dataset V2**
179
+ Developed by **Pyra Labs**