Skip to content

Commit

Permalink
feat: prompt_style applied to all LLMs + extra LLM params. (#1835)
Browse files Browse the repository at this point in the history
* Updated prompt_style to be moved to the main LLM setting since all LLMs from llama_index can utilize this.  I also included temperature, context window size, max_tokens, max_new_tokens into the openailike to help ensure the settings are consistent from the other implementations.

* Removed prompt_style from llamacpp entirely

* Fixed settings-local.yaml to include prompt_style in the LLM settings instead of llamacpp.
  • Loading branch information
icsy7867 authored Apr 30, 2024
1 parent c1802e7 commit e21bf20
Show file tree
Hide file tree
Showing 4 changed files with 22 additions and 18 deletions.
11 changes: 8 additions & 3 deletions private_gpt/components/llm/llm_component.py
Original file line number Diff line number Diff line change
Expand Up @@ -51,7 +51,7 @@ def __init__(self, settings: Settings) -> None:
"Local dependencies not found, install with `poetry install --extras llms-llama-cpp`"
) from e

prompt_style = get_prompt_style(settings.llamacpp.prompt_style)
prompt_style = get_prompt_style(settings.llm.prompt_style)
settings_kwargs = {
"tfs_z": settings.llamacpp.tfs_z, # ollama and llama-cpp
"top_k": settings.llamacpp.top_k, # ollama and llama-cpp
Expand Down Expand Up @@ -109,15 +109,20 @@ def __init__(self, settings: Settings) -> None:
raise ImportError(
"OpenAILike dependencies not found, install with `poetry install --extras llms-openai-like`"
) from e

prompt_style = get_prompt_style(settings.llm.prompt_style)
openai_settings = settings.openai
self.llm = OpenAILike(
api_base=openai_settings.api_base,
api_key=openai_settings.api_key,
model=openai_settings.model,
is_chat_model=True,
max_tokens=None,
max_tokens=settings.llm.max_new_tokens,
api_version="",
temperature=settings.llm.temperature,
context_window=settings.llm.context_window,
max_new_tokens=settings.llm.max_new_tokens,
messages_to_prompt=prompt_style.messages_to_prompt,
completion_to_prompt=prompt_style.completion_to_prompt,
)
case "ollama":
try:
Expand Down
23 changes: 11 additions & 12 deletions private_gpt/settings/settings.py
Original file line number Diff line number Diff line change
Expand Up @@ -104,6 +104,17 @@ class LLMSettings(BaseModel):
0.1,
description="The temperature of the model. Increasing the temperature will make the model answer more creatively. A value of 0.1 would be more factual.",
)
prompt_style: Literal["default", "llama2", "tag", "mistral", "chatml"] = Field(
"llama2",
description=(
"The prompt style to use for the chat engine. "
"If `default` - use the default prompt style from the llama_index. It should look like `role: message`.\n"
"If `llama2` - use the llama2 prompt style from the llama_index. Based on `<s>`, `[INST]` and `<<SYS>>`.\n"
"If `tag` - use the `tag` prompt style. It should look like `<|role|>: message`. \n"
"If `mistral` - use the `mistral prompt style. It shoudl look like <s>[INST] {System Prompt} [/INST]</s>[INST] { UserInstructions } [/INST]"
"`llama2` is the historic behaviour. `default` might work better with your custom models."
),
)


class VectorstoreSettings(BaseModel):
Expand All @@ -117,18 +128,6 @@ class NodeStoreSettings(BaseModel):
class LlamaCPPSettings(BaseModel):
llm_hf_repo_id: str
llm_hf_model_file: str
prompt_style: Literal["default", "llama2", "tag", "mistral", "chatml"] = Field(
"llama2",
description=(
"The prompt style to use for the chat engine. "
"If `default` - use the default prompt style from the llama_index. It should look like `role: message`.\n"
"If `llama2` - use the llama2 prompt style from the llama_index. Based on `<s>`, `[INST]` and `<<SYS>>`.\n"
"If `tag` - use the `tag` prompt style. It should look like `<|role|>: message`. \n"
"If `mistral` - use the `mistral prompt style. It shoudl look like <s>[INST] {System Prompt} [/INST]</s>[INST] { UserInstructions } [/INST]"
"`llama2` is the historic behaviour. `default` might work better with your custom models."
),
)

tfs_z: float = Field(
1.0,
description="Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting.",
Expand Down
4 changes: 2 additions & 2 deletions settings-local.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -8,9 +8,9 @@ llm:
max_new_tokens: 512
context_window: 3900
tokenizer: mistralai/Mistral-7B-Instruct-v0.2
prompt_style: "mistral"

llamacpp:
prompt_style: "mistral"
llm_hf_repo_id: TheBloke/Mistral-7B-Instruct-v0.2-GGUF
llm_hf_model_file: mistral-7b-instruct-v0.2.Q4_K_M.gguf

Expand All @@ -24,4 +24,4 @@ vectorstore:
database: qdrant

qdrant:
path: local_data/private_gpt/qdrant
path: local_data/private_gpt/qdrant
2 changes: 1 addition & 1 deletion settings.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,7 @@ ui:

llm:
mode: llamacpp
prompt_style: "mistral"
# Should be matching the selected model
max_new_tokens: 512
context_window: 3900
Expand All @@ -53,7 +54,6 @@ rag:
top_n: 1

llamacpp:
prompt_style: "mistral"
llm_hf_repo_id: TheBloke/Mistral-7B-Instruct-v0.2-GGUF
llm_hf_model_file: mistral-7b-instruct-v0.2.Q4_K_M.gguf
tfs_z: 1.0 # Tail free sampling is used to reduce the impact of less probable tokens from the output. A higher value (e.g., 2.0) will reduce the impact more, while a value of 1.0 disables this setting
Expand Down

0 comments on commit e21bf20

Please sign in to comment.