Fix prompt assembly for llama (#952)
Browse files* start at index 0
* add test to check for missing turns
* apply black
* Update test_prompt_tokenizers.py
* Update src/axolotl/monkeypatch/fastchat_conversation_turns.py
Co-authored-by: Motoki Wu <[email protected]>
* fix linting
* apply black
* add more tests for llama/sharegpt
* make logic clearer
---------
Co-authored-by: Motoki Wu <[email protected]>
src/axolotl/monkeypatch/fastchat_conversation_turns.py
CHANGED
@@ -83,14 +83,21 @@ def get_turns( # pylint: disable=too-many-return-statements
|
|
83 |
yield role + ":", ""
|
84 |
return
|
85 |
if self.sep_style == SeparatorStyle.LLAMA2:
|
86 |
-
seps = [self.sep, self.sep2]
|
87 |
if self.system_message:
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
yield "", system_prompt
|
89 |
-
|
90 |
-
yield "", "[INST] "
|
91 |
-
for i, (role, message) in enumerate(self.messages[1:]):
|
92 |
if message:
|
93 |
-
|
|
|
|
|
|
|
|
|
94 |
else:
|
95 |
yield role, ""
|
96 |
return
|
|
|
83 |
yield role + ":", ""
|
84 |
return
|
85 |
if self.sep_style == SeparatorStyle.LLAMA2:
|
|
|
86 |
if self.system_message:
|
87 |
+
if self.messages:
|
88 |
+
# For llama, the system message is incorporated into the first human instruction
|
89 |
+
first_role, first_msg = self.messages[0]
|
90 |
+
if first_role == self.roles[0]:
|
91 |
+
system_prompt += first_msg
|
92 |
+
self.messages.pop(0)
|
93 |
yield "", system_prompt
|
94 |
+
for i, (role, message) in enumerate(self.messages):
|
|
|
|
|
95 |
if message:
|
96 |
+
if (i % 2 == 0 and not self.system_message) or (
|
97 |
+
i % 2 != 0 and self.system_message
|
98 |
+
):
|
99 |
+
role = "<s> " + role
|
100 |
+
yield role + " ", message
|
101 |
else:
|
102 |
yield role, ""
|
103 |
return
|
tests/test_prompt_tokenizers.py
CHANGED
@@ -114,6 +114,76 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
|
|
114 |
in self._caplog.records[0].message
|
115 |
)
|
116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
def test_sharegpt_changes_roles(self):
|
118 |
conversation = {
|
119 |
"roles": ["USER", "CHARACTER"],
|
|
|
114 |
in self._caplog.records[0].message
|
115 |
)
|
116 |
|
117 |
+
def test_sharegpt_llama(self):
|
118 |
+
"Make sure the sharegpt/llama is tokenized and formatted correctly."
|
119 |
+
prompter = ShareGPTPrompterV2(conversation="llama-2")
|
120 |
+
strat = ShareGPTPromptTokenizingStrategy(
|
121 |
+
prompter,
|
122 |
+
self.tokenizer,
|
123 |
+
False,
|
124 |
+
2048,
|
125 |
+
)
|
126 |
+
|
127 |
+
def tokenize(conv):
|
128 |
+
return strat.tokenize_prompt(conv)["input_ids"]
|
129 |
+
|
130 |
+
def decode(ids):
|
131 |
+
return strat.tokenizer.decode(ids)
|
132 |
+
|
133 |
+
# Multi-turn conversations
|
134 |
+
multi_turn_conv = {
|
135 |
+
"conversations": [
|
136 |
+
{"from": "system", "value": "lorem"},
|
137 |
+
{"from": "human", "value": "abc"},
|
138 |
+
{"from": "gpt", "value": "ipsum"},
|
139 |
+
{"from": "human", "value": "123"},
|
140 |
+
{"from": "gpt", "value": "sit"},
|
141 |
+
]
|
142 |
+
}
|
143 |
+
# fmt: off
|
144 |
+
mt_ids = tokenize(multi_turn_conv)
|
145 |
+
assert decode(mt_ids) == '<s> [INST] <<SYS>>\nlorem\n<</SYS>>\n\nabc [/INST] ipsum</s><s> [INST] 123 [/INST] sit</s>'
|
146 |
+
assert mt_ids == [1, 518, 25580, 29962, 3532, 14816, 29903, 6778, 13, 29880, 3668, 13, 29966, 829, 14816, 29903, 6778, 13, 13, 10736, 518, 29914, 25580, 29962, 23421, 2, 1, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
|
147 |
+
|
148 |
+
# Single-turn conversations
|
149 |
+
single_turn_conv = {
|
150 |
+
"conversations": [
|
151 |
+
{"from": "system", "value": "lorem"},
|
152 |
+
{"from": "human", "value": "abc"},
|
153 |
+
{"from": "gpt", "value": "ipsum"},
|
154 |
+
]
|
155 |
+
}
|
156 |
+
|
157 |
+
st_ids = tokenize(single_turn_conv)
|
158 |
+
assert decode(st_ids) == '<s> [INST] <<SYS>>\nlorem\n<</SYS>>\n\nabc [/INST] ipsum</s>'
|
159 |
+
assert st_ids == [1, 518, 25580, 29962, 3532, 14816, 29903, 6778, 13, 29880, 3668, 13, 29966, 829, 14816, 29903, 6778, 13, 13, 10736, 518, 29914, 25580, 29962, 23421, 2]
|
160 |
+
|
161 |
+
# No system message, single-turn
|
162 |
+
no_sys_conv = {
|
163 |
+
"conversations": [
|
164 |
+
{"from": "human", "value": "abc"},
|
165 |
+
{"from": "gpt", "value": "ipsum"},
|
166 |
+
]
|
167 |
+
}
|
168 |
+
|
169 |
+
ns_ids = tokenize(no_sys_conv)
|
170 |
+
assert decode(ns_ids) == '<s> [INST] abc [/INST] ipsum</s>'
|
171 |
+
assert ns_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2]
|
172 |
+
|
173 |
+
# No system message, multi-turn
|
174 |
+
no_sys_mt_conv = {
|
175 |
+
"conversations": [
|
176 |
+
{"from": "human", "value": "abc"},
|
177 |
+
{"from": "gpt", "value": "ipsum"},
|
178 |
+
{"from": "human", "value": "123"},
|
179 |
+
{"from": "gpt", "value": "sit"},
|
180 |
+
]
|
181 |
+
}
|
182 |
+
ns_mt_ids = tokenize(no_sys_mt_conv)
|
183 |
+
assert decode(ns_mt_ids) == '<s> [INST] abc [/INST] ipsum</s><s> [INST] 123 [/INST] sit</s>'
|
184 |
+
assert ns_mt_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2, 1, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
|
185 |
+
# fmt: on
|
186 |
+
|
187 |
def test_sharegpt_changes_roles(self):
|
188 |
conversation = {
|
189 |
"roles": ["USER", "CHARACTER"],
|