File size: 3,499 Bytes
1eaf4ea
7b62db0
 
 
 
1eaf4ea
7b62db0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language:
- en
tags:
- llama
---

# OpenChat: Less is More for Open-source Models

OpenChat is a series of open-source language models fine-tuned on very little diverse and high-quality multi-round conversations. The [dataset](https://huggingface.co/datasets/openchat/openchat_sharegpt4_dataset) contains only ~6K GPT-4 conversations filtered from the 90K ShareGPT conversations.

Generic models:

 - OpenChat: based on LLaMA-13B (2048 context length)
   - **105.7%** of ChatGPT score on Vicuna GPT-4 evaluation
   - **80.87%** Win-rate on AlpacaEval
   - **🚀 Only used 6K data for finetuning!!!**
 - OpenChat-8192: based on LLaMA-13B  (extended to 8192 context length)
   - **106.6%** of ChatGPT score on Vicuna GPT-4 evaluation

Code models:

 - OpenCoderPlus: based on StarCoderPlus (native 8192 context length)
   - **102.5%** of ChatGPT score on Vicuna GPT-4 evaluation
   - **78.70%** Win-rate on AlpacaEval

**NOTE:** Please load the pretrained models using *bfloat16*

## Conversation Template

The conversation template **involves concatenating tokens**.

Besides base model vocabulary, an end-of-turn token `<|end_of_turn|>` is added, with id `eot_token_id`.

```python
# OpenChat
[bos_token_id] + tokenize("Human: ") + tokenize(user_question) + [eot_token_id] + tokenize("Assistant: ")
# OpenCoder
tokenize("User:") + tokenize(user_question) + [eot_token_id] + tokenize("Assistant:")
```

*Hint: In BPE, `tokenize(A) + tokenize(B)` does not always equals to `tokenize(A + B)`*

Following is the code for generating the conversation templates:

```python
@dataclass
class ModelConfig:
    # Prompt
    system: Optional[str]

    role_prefix: dict
    ai_role: str
    eot_token: str
    bos_token: Optional[str] = None

    # Get template
    def generate_conversation_template(self, tokenize_fn, tokenize_special_fn, message_list):
        tokens = []
        masks = []

        # begin of sentence (bos)
        if self.bos_token:
            t = tokenize_special_fn(self.bos_token)
            tokens.append(t)
            masks.append(False)

        # System
        if self.system:
            t = tokenize_fn(self.system) + [tokenize_special_fn(self.eot_token)]
            tokens.extend(t)
            masks.extend([False] * len(t))

        # Messages
        for idx, message in enumerate(message_list):
            # Prefix
            t = tokenize_fn(self.role_prefix[message["from"]])
            tokens.extend(t)
            masks.extend([False] * len(t))

            # Message
            if "value" in message:
                t = tokenize_fn(message["value"]) + [tokenize_special_fn(self.eot_token)]
                tokens.extend(t)
                masks.extend([message["from"] == self.ai_role] * len(t))
            else:
                assert idx == len(message_list) - 1, "Empty message for completion must be on the last."

        return tokens, masks


MODEL_CONFIG_MAP = {
    # OpenChat / OpenChat-8192
    "openchat": ModelConfig(
        # Prompt
        system=None,

        role_prefix={
            "human": "Human: ",
            "gpt": "Assistant: "
        },
        ai_role="gpt",
        eot_token="<|end_of_turn|>",
        bos_token="<s>",
    ),

    # OpenCoder / OpenCoderPlus
    "opencoder": ModelConfig(
        # Prompt
        system=None,

        role_prefix={
            "human": "User:",
            "gpt": "Assistant:"
        },
        ai_role="gpt",
        eot_token="<|end_of_turn|>",
        bos_token=None,
    )
}
```