fix camel ai, add guanaco/oasst mapping for sharegpt
Browse files
src/axolotl/prompt_strategies/alpaca_chat.py
CHANGED
|
@@ -49,7 +49,7 @@ class CamelAIPromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
|
|
| 49 |
return (
|
| 50 |
prompt["message_1"],
|
| 51 |
"",
|
| 52 |
-
prompt["
|
| 53 |
)
|
| 54 |
|
| 55 |
|
|
|
|
| 49 |
return (
|
| 50 |
prompt["message_1"],
|
| 51 |
"",
|
| 52 |
+
prompt["message_2"],
|
| 53 |
)
|
| 54 |
|
| 55 |
|
src/axolotl/prompt_strategies/sharegpt_simple.py
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Module containing the SimpleShareGPTPromptTokenizingStrategy class"""
|
| 2 |
+
|
| 3 |
+
from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
|
| 4 |
+
from axolotl.prompters import PromptStyle, ShareGPTPrompter
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def load(tokenizer, cfg):
|
| 8 |
+
return SimpleShareGPTPromptTokenizingStrategy(
|
| 9 |
+
ShareGPTPrompter(PromptStyle.CHAT.value),
|
| 10 |
+
tokenizer,
|
| 11 |
+
cfg.train_on_inputs,
|
| 12 |
+
cfg.sequence_len,
|
| 13 |
+
)
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def load_guanaco(tokenizer, cfg):
|
| 17 |
+
return GuanacoShareGPTPromptTokenizingStrategy(
|
| 18 |
+
ShareGPTPrompter(PromptStyle.CHAT.value),
|
| 19 |
+
tokenizer,
|
| 20 |
+
cfg.train_on_inputs,
|
| 21 |
+
cfg.sequence_len,
|
| 22 |
+
)
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
|
| 26 |
+
"""
|
| 27 |
+
basic sharegpt strategy to grab conversations from the sample row
|
| 28 |
+
"""
|
| 29 |
+
|
| 30 |
+
def get_conversation_thread(self, prompt):
|
| 31 |
+
return prompt["conversations"]
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
|
| 35 |
+
"""
|
| 36 |
+
sharegpt strategy that remaps oasst data to sharegpt format
|
| 37 |
+
"""
|
| 38 |
+
|
| 39 |
+
def get_conversation_thread(self, prompt):
|
| 40 |
+
conversations = prompt["conversations"]
|
| 41 |
+
# remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
|
| 42 |
+
role_map = {"prompter": "human", "assistant": "gpt"}
|
| 43 |
+
turns = [
|
| 44 |
+
{"from": role_map[t["role"]], "value": t["text"]} for t in conversations
|
| 45 |
+
]
|
| 46 |
+
return turns
|