File size: 6,597 Bytes
59bb219
b7d8a7d
40a88e8
e7d3e2d
 
 
59bb219
 
e7d3e2d
b7d8a7d
 
 
 
e7d3e2d
40a88e8
 
98b4762
 
 
 
 
 
 
 
 
 
 
 
e7d3e2d
b7d8a7d
 
 
 
 
 
 
 
 
 
59bb219
 
e7d3e2d
 
 
 
 
 
40a88e8
0800885
e7d3e2d
 
 
 
40a88e8
e7d3e2d
59bb219
 
 
 
0800885
 
 
59bb219
 
ba043a3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aac4b76
 
e7d3e2d
aac4b76
 
 
 
 
 
59bb219
 
e7d3e2d
59bb219
 
 
 
 
 
b7d8a7d
 
 
 
 
 
 
 
 
 
 
 
 
 
59bb219
 
 
 
 
2598c9f
0800885
 
 
 
 
 
 
 
 
59bb219
0800885
 
 
2598c9f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0800885
40a88e8
 
 
 
 
 
2598c9f
0800885
 
59bb219
 
aac4b76
 
 
 
 
 
 
 
 
 
 
 
59bb219
 
 
 
 
 
 
 
 
 
 
 
 
ba043a3
 
 
 
 
 
 
 
 
 
 
 
 
 
b7d8a7d
 
 
 
 
 
 
 
 
 
 
 
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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
"""Module containing the SimpleShareGPTPromptTokenizingStrategy class"""

import logging
from typing import Any, Dict, Optional

from fastchat.conversation import Conversation, SeparatorStyle, register_conv_template

from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
from axolotl.prompters import ShareGPTPrompterV2
from axolotl.utils.tokenization import (
    chatml_to_conversation,
    merge_consecutive_messages,
)

LOG = logging.getLogger("axolotl")


def register_chatml_template(system_message=None):
    system_message = system_message or "You are a helpful assistant."
    register_conv_template(
        Conversation(
            name="chatml",
            system_template="<|im_start|>system\n{system_message}",
            system_message=system_message,
            roles=["<|im_start|>user", "<|im_start|>assistant"],
            sep_style=SeparatorStyle.CHATML,
            sep="<|im_end|>",
        )
    )
    register_conv_template(
        Conversation(
            name="chatml_glaive",
            system_template="<|im_start|>system\n{system_message}",
            system_message=system_message,
            roles=["<|im_start|>user", "<|im_start|>assistant", "<|im_start|>tool"],
            sep_style=SeparatorStyle.CHATML,
            sep="<|im_end|>",
        )
    )


def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
    conversation = (
        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else None
    )
    field_human = ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
    field_model = ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
    roles = ds_cfg["roles"].to_dict() if ds_cfg and "roles" in ds_cfg else None
    strategy = SimpleShareGPTPromptTokenizingStrategy(
        ShareGPTPrompterV2(
            conversation=conversation,
            role_key_model=field_model,
            role_key_human=field_human,
            roles=roles,
        ),
        tokenizer,
        cfg.train_on_inputs,
        cfg.sequence_len,
    )
    if ds_cfg and "strict" in ds_cfg:
        strategy.strict = ds_cfg["strict"]
    return strategy


def load_ultrachat(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
    conversation = (
        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else None
    )
    strategy = UltrachatShareGPTPromptTokenizingStrategy(
        ShareGPTPrompterV2(
            conversation=conversation,
        ),
        tokenizer,
        cfg.train_on_inputs,
        cfg.sequence_len,
    )
    if ds_cfg and "strict" in ds_cfg:
        strategy.strict = ds_cfg["strict"]
    return strategy


def load_role(tokenizer, cfg):
    return SimpleRoleShareGPTPromptTokenizingStrategy(
        ShareGPTPrompterV2(),
        tokenizer,
        cfg.train_on_inputs,
        cfg.sequence_len,
    )


def load_guanaco(tokenizer, cfg):
    return GuanacoShareGPTPromptTokenizingStrategy(
        ShareGPTPrompterV2(),
        tokenizer,
        cfg.train_on_inputs,
        cfg.sequence_len,
    )


def load_glaive(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
    conversation = (
        ds_cfg["conversation"]
        if ds_cfg and "conversation" in ds_cfg
        else "chatml_glaive"
    )
    return GlaiveShareGPTPromptTokenizingStrategy(
        ShareGPTPrompterV2(conversation=conversation),
        tokenizer,
        cfg.train_on_inputs,
        cfg.sequence_len,
    )


class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
    """
    basic sharegpt strategy to grab conversations from the sample row
    """

    _strict = False

    @property
    def strict(self):
        return self._strict

    @strict.setter
    def strict(self, strict):
        self._strict = strict

    def get_conversation_thread(self, prompt):
        conversations = prompt["conversations"]
        if self.strict:
            return conversations
        role_key = "from"
        if "role" in conversations[0].keys():
            role_key = "role"
        value_key = "value"
        if "text" in conversations[0].keys():
            value_key = "text"
        elif "content" in conversations[0].keys():
            value_key = "content"
        # remap roles - allow for assistant turn"
        role_map = {
            "user": "human",
            "human": "human",
            "assistant": "gpt",
            "gpt": "gpt",
            "system": "system",
        }
        turns = [
            {
                "from": (
                    role_map[t[role_key]] if t[role_key] in role_map else t[role_key]
                ),
                "value": t[value_key],
            }
            for t in conversations
        ]
        return turns


class SimpleRoleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
    """
    basic sharegpt strategy to grab conversations from the sample row, but uses role instead of from
    """

    def get_conversation_thread(self, prompt):
        conversations = prompt["conversations"]
        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
        turns = [{"from": t["role"], "value": t["value"]} for t in conversations]
        return turns


class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
    """
    sharegpt strategy that remaps oasst data to sharegpt format
    """

    def get_conversation_thread(self, prompt):
        conversations = prompt["conversations"]
        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
        role_map = {"prompter": "human", "assistant": "gpt"}
        turns = [
            {"from": role_map[t["role"]], "value": t["text"]} for t in conversations
        ]
        return turns


class UltrachatShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
    """
    sharegpt strategy that remaps ultrachat data to sharegpt format
    """

    def get_conversation_thread(self, prompt):
        conversations = prompt["messages"]
        role_map = {"user": "human", "assistant": "gpt"}
        turns = [
            {"from": role_map[t["role"]], "value": t["content"]} for t in conversations
        ]
        return turns


class GlaiveShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
    """
    sharegpt strategy that remaps glaive data to sharegpt format
    """

    def get_conversation_thread(self, prompt):
        conversation = chatml_to_conversation(prompt)
        conversation = merge_consecutive_messages(conversation)

        return conversation