phoenix-oss/llama-stack
1
0
Fork 0
forked from phoenix-oss/llama-stack-mirror
Code Pull requests Releases Packages 2 Activity Actions

feat(api): (1/n) datasets api clean up (#1573)

Browse source 
## PR Stack
- https://github.com/meta-llama/llama-stack/pull/1573
- https://github.com/meta-llama/llama-stack/pull/1625
- https://github.com/meta-llama/llama-stack/pull/1656
- https://github.com/meta-llama/llama-stack/pull/1657
- https://github.com/meta-llama/llama-stack/pull/1658
- https://github.com/meta-llama/llama-stack/pull/1659
- https://github.com/meta-llama/llama-stack/pull/1660

**Client SDK**
- https://github.com/meta-llama/llama-stack-client-python/pull/203

**CI**
- 1391130488
<img width="1042" alt="image"
src="https://github.com/user-attachments/assets/69636067-376d-436b-9204-896e2dd490ca"
/>
-- the test_rag_agent_with_attachments is flaky and not related to this
PR

## Doc
<img width="789" alt="image"
src="https://github.com/user-attachments/assets/b88390f3-73d6-4483-b09a-a192064e32d9"
/>


## Client Usage
```python
client.datasets.register(
    source={
        "type": "uri",
        "uri": "lsfs://mydata.jsonl",
    },
    schema="jsonl_messages",
    # optional 
    dataset_id="my_first_train_data"
)

# quick prototype debugging
client.datasets.register(
    data_reference={
        "type": "rows",
        "rows": [
                "messages": [...],
        ],
    },
    schema="jsonl_messages",
)
```

## Test Plan
- CI:
1387805545

```
LLAMA_STACK_CONFIG=fireworks pytest -v tests/integration/datasets/test_datasets.py
```

```
LLAMA_STACK_CONFIG=fireworks pytest -v tests/integration/scoring/test_scoring.py
```

```
pytest -v -s --nbval-lax ./docs/notebooks/Llama_Stack_Benchmark_Evals.ipynb
```
This commit is contained in:
Xi Yan 2025-03-17 16:55:45 -07:00 • committed by GitHub
parent 3b35a39b8b
commit 5287b437ae
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
29 changed files with 2593 additions and 2296 deletions
Download patch file Download diff file Expand all files Collapse all files

34
llama_stack/apis/datasetio/datasetio.py
View file

@ -13,19 +13,16 @@ from llama_stack.schema_utils import json_schema_type, webmethod
@json_schema_type
class PaginatedRowsResult(BaseModel):
class IterrowsResponse(BaseModel):
"""
A paginated list of rows from a dataset.
:param rows: The rows in the current page.
:param total_count: The total number of rows in the dataset.
:param next_page_token: The token to get the next page of rows.
:param data: The rows in the current page.
:param next_start_index: Index into dataset for the first row in the next page. None if there are no more rows.
"""
# the rows obey the DatasetSchema for the given dataset
rows: List[Dict[str, Any]]
total_count: int
next_page_token: Optional[str] = None
data: List[Dict[str, Any]]
next_start_index: Optional[int] = None
class DatasetStore(Protocol):
@ -37,22 +34,21 @@ class DatasetIO(Protocol):
# keeping for aligning with inference/safety, but this is not used
dataset_store: DatasetStore
@webmethod(route="/datasetio/rows", method="GET")
async def get_rows_paginated(
# TODO(xiyan): there's a flakiness here where setting route to "/datasets/" here will not result in proper routing
@webmethod(route="/datasetio/iterrows/{dataset_id:path}", method="GET")
async def iterrows(
self,
dataset_id: str,
rows_in_page: int,
page_token: Optional[str] = None,
filter_condition: Optional[str] = None,
) -> PaginatedRowsResult:
"""Get a paginated list of rows from a dataset.
start_index: Optional[int] = None,
limit: Optional[int] = None,
) -> IterrowsResponse:
"""Get a paginated list of rows from a dataset. Uses cursor-based pagination.
:param dataset_id: The ID of the dataset to get the rows from.
:param rows_in_page: The number of rows to get per page.
:param page_token: The token to get the next page of rows.
:param filter_condition: (Optional) A condition to filter the rows by.
:param start_index: Index into dataset for the first row to get. Get all rows if None.
:param limit: The number of rows to get.
"""
...
@webmethod(route="/datasetio/rows", method="POST")
@webmethod(route="/datasetio/append-rows/{dataset_id:path}", method="POST")
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None: ...

163
llama_stack/apis/datasets/datasets.py
View file

@ -4,19 +4,102 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Literal, Optional, Protocol
from enum import Enum
from typing import Annotated, Any, Dict, List, Literal, Optional, Protocol, Union
from pydantic import BaseModel, Field
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.common.type_system import ParamType
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.schema_utils import json_schema_type, webmethod
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
class DatasetPurpose(str, Enum):
"""
Purpose of the dataset. Each purpose has a required input data schema.
:cvar post-training/messages: The dataset contains messages used for post-training.
{
"messages": [
{"role": "user", "content": "Hello, world!"},
{"role": "assistant", "content": "Hello, world!"},
]
}
:cvar eval/question-answer: The dataset contains a question column and an answer column.
{
"question": "What is the capital of France?",
"answer": "Paris"
}
:cvar eval/messages-answer: The dataset contains a messages column with list of messages and an answer column.
{
"messages": [
{"role": "user", "content": "Hello, my name is John Doe."},
{"role": "assistant", "content": "Hello, John Doe. How can I help you today?"},
{"role": "user", "content": "What's my name?"},
],
"answer": "John Doe"
}
"""
post_training_messages = "post-training/messages"
eval_question_answer = "eval/question-answer"
eval_messages_answer = "eval/messages-answer"
# TODO: add more schemas here
class DatasetType(Enum):
"""
Type of the dataset source.
:cvar uri: The dataset can be obtained from a URI.
:cvar rows: The dataset is stored in rows.
"""
uri = "uri"
rows = "rows"
@json_schema_type
class URIDataSource(BaseModel):
"""A dataset that can be obtained from a URI.
:param uri: The dataset can be obtained from a URI. E.g.
- "https://mywebsite.com/mydata.jsonl"
- "lsfs://mydata.jsonl"
- "data:csv;base64,{base64_content}"
"""
type: Literal["uri"] = "uri"
uri: str
@json_schema_type
class RowsDataSource(BaseModel):
"""A dataset stored in rows.
:param rows: The dataset is stored in rows. E.g.
- [
{"messages": [{"role": "user", "content": "Hello, world!"}, {"role": "assistant", "content": "Hello, world!"}]}
]
"""
type: Literal["rows"] = "rows"
rows: List[Dict[str, Any]]
DataSource = register_schema(
Annotated[
Union[URIDataSource, RowsDataSource],
Field(discriminator="type"),
],
name="DataSource",
)
class CommonDatasetFields(BaseModel):
dataset_schema: Dict[str, ParamType]
url: URL
"""
Common fields for a dataset.
"""
purpose: DatasetPurpose
source: DataSource
metadata: Dict[str, Any] = Field(
default_factory=dict,
description="Any additional metadata for this dataset",
@ -50,13 +133,69 @@ class Datasets(Protocol):
@webmethod(route="/datasets", method="POST")
async def register_dataset(
self,
dataset_id: str,
dataset_schema: Dict[str, ParamType],
url: URL,
provider_dataset_id: Optional[str] = None,
provider_id: Optional[str] = None,
purpose: DatasetPurpose,
source: DataSource,
metadata: Optional[Dict[str, Any]] = None,
) -> None: ...
dataset_id: Optional[str] = None,
) -> Dataset:
"""
Register a new dataset.
:param purpose: The purpose of the dataset. One of
- "post-training/messages": The dataset contains a messages column with list of messages for post-training.
{
"messages": [
{"role": "user", "content": "Hello, world!"},
{"role": "assistant", "content": "Hello, world!"},
]
}
- "eval/question-answer": The dataset contains a question column and an answer column for evaluation.
{
"question": "What is the capital of France?",
"answer": "Paris"
}
- "eval/messages-answer": The dataset contains a messages column with list of messages and an answer column for evaluation.
{
"messages": [
{"role": "user", "content": "Hello, my name is John Doe."},
{"role": "assistant", "content": "Hello, John Doe. How can I help you today?"},
{"role": "user", "content": "What's my name?"},
],
"answer": "John Doe"
}
:param source: The data source of the dataset. Ensure that the data source schema is compatible with the purpose of the dataset. Examples:
- {
"type": "uri",
"uri": "https://mywebsite.com/mydata.jsonl"
}
- {
"type": "uri",
"uri": "lsfs://mydata.jsonl"
}
- {
"type": "uri",
"uri": "data:csv;base64,{base64_content}"
}
- {
"type": "uri",
"uri": "huggingface://llamastack/simpleqa?split=train"
}
- {
"type": "rows",
"rows": [
{
"messages": [
{"role": "user", "content": "Hello, world!"},
{"role": "assistant", "content": "Hello, world!"},
]
}
]
}
:param metadata: The metadata for the dataset.
- E.g. {"description": "My dataset"}
:param dataset_id: The ID of the dataset. If not provided, an ID will be generated.
"""
...
@webmethod(route="/datasets/{dataset_id:path}", method="GET")
async def get_dataset(

51
llama_stack/distribution/routers/routers.py
View file

@ -12,7 +12,8 @@ from llama_stack.apis.common.content_types import (
InterleavedContent,
InterleavedContentItem,
)
from llama_stack.apis.datasetio import DatasetIO, PaginatedRowsResult
from llama_stack.apis.datasetio import DatasetIO, IterrowsResponse
from llama_stack.apis.datasets import DatasetPurpose, DataSource
from llama_stack.apis.eval import (
BenchmarkConfig,
Eval,
@ -160,7 +161,11 @@ class InferenceRouter(Inference):
await self.routing_table.register_model(model_id, provider_model_id, provider_id, metadata, model_type)
def _construct_metrics(
self, prompt_tokens: int, completion_tokens: int, total_tokens: int, model: Model
self,
prompt_tokens: int,
completion_tokens: int,
total_tokens: int,
model: Model,
) -> List[MetricEvent]:
"""Constructs a list of MetricEvent objects containing token usage metrics.
@ -298,7 +303,12 @@ class InferenceRouter(Inference):
completion_text += chunk.event.delta.text
if chunk.event.event_type == ChatCompletionResponseEventType.complete:
completion_tokens = await self._count_tokens(
[CompletionMessage(content=completion_text, stop_reason=StopReason.end_of_turn)],
[
CompletionMessage(
content=completion_text,
stop_reason=StopReason.end_of_turn,
)
],
tool_config.tool_prompt_format,
)
total_tokens = (prompt_tokens or 0) + (completion_tokens or 0)
@ -471,21 +481,36 @@ class DatasetIORouter(DatasetIO):
logger.debug("DatasetIORouter.shutdown")
pass
async def get_rows_paginated(
async def register_dataset(
self,
purpose: DatasetPurpose,
source: DataSource,
metadata: Optional[Dict[str, Any]] = None,
dataset_id: Optional[str] = None,
) -> None:
logger.debug(
f"DatasetIORouter.register_dataset: {purpose=} {source=} {metadata=} {dataset_id=}",
)
await self.routing_table.register_dataset(
purpose=purpose,
source=source,
metadata=metadata,
dataset_id=dataset_id,
)
async def iterrows(
self,
dataset_id: str,
rows_in_page: int,
page_token: Optional[str] = None,
filter_condition: Optional[str] = None,
) -> PaginatedRowsResult:
start_index: Optional[int] = None,
limit: Optional[int] = None,
) -> IterrowsResponse:
logger.debug(
f"DatasetIORouter.get_rows_paginated: {dataset_id}, rows_in_page={rows_in_page}",
f"DatasetIORouter.iterrows: {dataset_id}, {start_index=} {limit=}",
)
return await self.routing_table.get_provider_impl(dataset_id).get_rows_paginated(
return await self.routing_table.get_provider_impl(dataset_id).iterrows(
dataset_id=dataset_id,
rows_in_page=rows_in_page,
page_token=page_token,
filter_condition=filter_condition,
start_index=start_index,
limit=limit,
)
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None:

52
llama_stack/distribution/routers/routing_tables.py
View file

@ -5,6 +5,7 @@
# the root directory of this source tree.
import logging
import uuid
from typing import Any, Dict, List, Optional
from pydantic import TypeAdapter
@ -12,7 +13,14 @@ from pydantic import TypeAdapter
from llama_stack.apis.benchmarks import Benchmark, Benchmarks, ListBenchmarksResponse
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.common.type_system import ParamType
from llama_stack.apis.datasets import Dataset, Datasets, ListDatasetsResponse
from llama_stack.apis.datasets import (
Dataset,
DatasetPurpose,
Datasets,
DatasetType,
DataSource,
ListDatasetsResponse,
)
from llama_stack.apis.models import ListModelsResponse, Model, Models, ModelType
from llama_stack.apis.resource import ResourceType
from llama_stack.apis.scoring_functions import (
@ -352,34 +360,42 @@ class DatasetsRoutingTable(CommonRoutingTableImpl, Datasets):
async def register_dataset(
self,
dataset_id: str,
dataset_schema: Dict[str, ParamType],
url: URL,
provider_dataset_id: Optional[str] = None,
provider_id: Optional[str] = None,
purpose: DatasetPurpose,
source: DataSource,
metadata: Optional[Dict[str, Any]] = None,
) -> None:
if provider_dataset_id is None:
provider_dataset_id = dataset_id
if provider_id is None:
# If provider_id not specified, use the only provider if it supports this dataset
if len(self.impls_by_provider_id) == 1:
provider_id = list(self.impls_by_provider_id.keys())[0]
dataset_id: Optional[str] = None,
) -> Dataset:
if not dataset_id:
dataset_id = f"dataset-{str(uuid.uuid4())}"
provider_dataset_id = dataset_id
# infer provider from source
if source.type == DatasetType.rows.value:
provider_id = "localfs"
elif source.type == DatasetType.uri.value:
# infer provider from uri
if source.uri.startswith("huggingface"):
provider_id = "huggingface"
else:
raise ValueError(
f"No provider specified and multiple providers available. Please specify a provider_id. Available providers: {self.impls_by_provider_id.keys()}"
)
provider_id = "localfs"
else:
raise ValueError(f"Unknown data source type: {source.type}")
if metadata is None:
metadata = {}
dataset = Dataset(
identifier=dataset_id,
provider_resource_id=provider_dataset_id,
provider_id=provider_id,
dataset_schema=dataset_schema,
url=url,
purpose=purpose,
source=source,
metadata=metadata,
)
await self.register_object(dataset)
return dataset
async def unregister_dataset(self, dataset_id: str) -> None:
dataset = await self.get_dataset(dataset_id)

7
llama_stack/distribution/ui/page/evaluations/native_eval.py
View file

@ -166,11 +166,10 @@ def run_evaluation_3():
eval_candidate = st.session_state["eval_candidate"]
dataset_id = benchmarks[selected_benchmark].dataset_id
rows = llama_stack_api.client.datasetio.get_rows_paginated(
rows = llama_stack_api.client.datasets.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
)
total_rows = len(rows.rows)
total_rows = len(rows.data)
# Add number of examples control
num_rows = st.number_input(
"Number of Examples to Evaluate",
@ -195,7 +194,7 @@ def run_evaluation_3():
if st.button("Run Evaluation"):
progress_text = "Running evaluation..."
progress_bar = st.progress(0, text=progress_text)
rows = rows.rows
rows = rows.data
if num_rows < total_rows:
rows = rows[:num_rows]

139
llama_stack/providers/inline/datasetio/localfs/datasetio.py
View file

@ -3,20 +3,14 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import base64
import os
from abc import ABC, abstractmethod
from dataclasses import dataclass
from typing import Any, Dict, List, Optional
from urllib.parse import urlparse
import pandas
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.datasetio import DatasetIO, PaginatedRowsResult
from llama_stack.apis.datasetio import DatasetIO, IterrowsResponse
from llama_stack.apis.datasets import Dataset
from llama_stack.providers.datatypes import DatasetsProtocolPrivate
from llama_stack.providers.utils.datasetio.url_utils import get_dataframe_from_url
from llama_stack.providers.utils.datasetio.url_utils import get_dataframe_from_uri
from llama_stack.providers.utils.kvstore import kvstore_impl
from .config import LocalFSDatasetIOConfig
@ -24,30 +18,7 @@ from .config import LocalFSDatasetIOConfig
DATASETS_PREFIX = "localfs_datasets:"
class BaseDataset(ABC):
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
@abstractmethod
def __len__(self) -> int:
raise NotImplementedError()
@abstractmethod
def __getitem__(self, idx):
raise NotImplementedError()
@abstractmethod
def load(self):
raise NotImplementedError()
@dataclass
class DatasetInfo:
dataset_def: Dataset
dataset_impl: BaseDataset
class PandasDataframeDataset(BaseDataset):
class PandasDataframeDataset:
def __init__(self, dataset_def: Dataset, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self.dataset_def = dataset_def
@ -64,23 +35,19 @@ class PandasDataframeDataset(BaseDataset):
else:
return self.df.iloc[idx].to_dict()
def _validate_dataset_schema(self, df) -> pandas.DataFrame:
# note that we will drop any columns in dataset that are not in the schema
df = df[self.dataset_def.dataset_schema.keys()]
# check all columns in dataset schema are present
assert len(df.columns) == len(self.dataset_def.dataset_schema)
# TODO: type checking against column types in dataset schema
return df
def load(self) -> None:
if self.df is not None:
return
df = get_dataframe_from_url(self.dataset_def.url)
if df is None:
raise ValueError(f"Failed to load dataset from {self.dataset_def.url}")
if self.dataset_def.source.type == "uri":
self.df = get_dataframe_from_uri(self.dataset_def.source.uri)
elif self.dataset_def.source.type == "rows":
self.df = pandas.DataFrame(self.dataset_def.source.rows)
else:
raise ValueError(f"Unsupported dataset source type: {self.dataset_def.source.type}")
self.df = self._validate_dataset_schema(df)
if self.df is None:
raise ValueError(f"Failed to load dataset from {self.dataset_def.url}")
class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
@ -99,95 +66,55 @@ class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
for dataset in stored_datasets:
dataset = Dataset.model_validate_json(dataset)
dataset_impl = PandasDataframeDataset(dataset)
self.dataset_infos[dataset.identifier] = DatasetInfo(
dataset_def=dataset,
dataset_impl=dataset_impl,
)
self.dataset_infos[dataset.identifier] = dataset
async def shutdown(self) -> None: ...
async def register_dataset(
self,
dataset: Dataset,
dataset_def: Dataset,
) -> None:
# Store in kvstore
key = f"{DATASETS_PREFIX}{dataset.identifier}"
key = f"{DATASETS_PREFIX}{dataset_def.identifier}"
await self.kvstore.set(
key=key,
value=dataset.json(),
)
dataset_impl = PandasDataframeDataset(dataset)
self.dataset_infos[dataset.identifier] = DatasetInfo(
dataset_def=dataset,
dataset_impl=dataset_impl,
value=dataset_def.model_dump_json(),
)
self.dataset_infos[dataset_def.identifier] = dataset_def
async def unregister_dataset(self, dataset_id: str) -> None:
key = f"{DATASETS_PREFIX}{dataset_id}"
await self.kvstore.delete(key=key)
del self.dataset_infos[dataset_id]
async def get_rows_paginated(
async def iterrows(
self,
dataset_id: str,
rows_in_page: int,
page_token: Optional[str] = None,
filter_condition: Optional[str] = None,
) -> PaginatedRowsResult:
dataset_info = self.dataset_infos.get(dataset_id)
dataset_info.dataset_impl.load()
start_index: Optional[int] = None,
limit: Optional[int] = None,
) -> IterrowsResponse:
dataset_def = self.dataset_infos[dataset_id]
dataset_impl = PandasDataframeDataset(dataset_def)
dataset_impl.load()
if page_token and not page_token.isnumeric():
raise ValueError("Invalid page_token")
start_index = start_index or 0
if page_token is None or len(page_token) == 0:
next_page_token = 0
if limit is None or limit == -1:
end = len(dataset_impl)
else:
next_page_token = int(page_token)
end = min(start_index + limit, len(dataset_impl))
start = next_page_token
if rows_in_page == -1:
end = len(dataset_info.dataset_impl)
else:
end = min(start + rows_in_page, len(dataset_info.dataset_impl))
rows = dataset_impl[start_index:end]
rows = dataset_info.dataset_impl[start:end]
return PaginatedRowsResult(
rows=rows,
total_count=len(rows),
next_page_token=str(end),
return IterrowsResponse(
data=rows,
next_start_index=end if end < len(dataset_impl) else None,
)
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None:
dataset_info = self.dataset_infos.get(dataset_id)
if dataset_info is None:
raise ValueError(f"Dataset with id {dataset_id} not found")
dataset_impl = dataset_info.dataset_impl
dataset_def = self.dataset_infos[dataset_id]
dataset_impl = PandasDataframeDataset(dataset_def)
dataset_impl.load()
new_rows_df = pandas.DataFrame(rows)
new_rows_df = dataset_impl._validate_dataset_schema(new_rows_df)
dataset_impl.df = pandas.concat([dataset_impl.df, new_rows_df], ignore_index=True)
url = str(dataset_info.dataset_def.url.uri)
parsed_url = urlparse(url)
if parsed_url.scheme == "file" or not parsed_url.scheme:
file_path = parsed_url.path
os.makedirs(os.path.dirname(file_path), exist_ok=True)
dataset_impl.df.to_csv(file_path, index=False)
elif parsed_url.scheme == "data":
# For data URLs, we need to update the base64-encoded content
if not parsed_url.path.startswith("text/csv;base64,"):
raise ValueError("Data URL must be a base64-encoded CSV")
csv_buffer = dataset_impl.df.to_csv(index=False)
base64_content = base64.b64encode(csv_buffer.encode("utf-8")).decode("utf-8")
dataset_info.dataset_def.url = URL(uri=f"data:text/csv;base64,{base64_content}")
else:
raise ValueError(
f"Unsupported URL scheme: {parsed_url.scheme}. Only file:// and data: URLs are supported for writing."
)

19
llama_stack/providers/inline/eval/meta_reference/eval.py
View file

@ -14,16 +14,11 @@ from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Datasets
from llama_stack.apis.inference import Inference, SystemMessage, UserMessage
from llama_stack.apis.scoring import Scoring
from llama_stack.distribution.datatypes import Api
from llama_stack.providers.datatypes import BenchmarksProtocolPrivate
from llama_stack.providers.inline.agents.meta_reference.agent_instance import (
MEMORY_QUERY_TOOL,
)
from llama_stack.providers.utils.common.data_schema_validator import (
ColumnName,
get_valid_schemas,
validate_dataset_schema,
)
from llama_stack.providers.utils.common.data_schema_validator import ColumnName
from llama_stack.providers.utils.kvstore import kvstore_impl
from .....apis.common.job_types import Job
@ -88,15 +83,17 @@ class MetaReferenceEvalImpl(
task_def = self.benchmarks[benchmark_id]
dataset_id = task_def.dataset_id
scoring_functions = task_def.scoring_functions
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.eval.value))
all_rows = await self.datasetio_api.get_rows_paginated(
# TODO (xiyan): validate dataset schema
# dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=(-1 if benchmark_config.num_examples is None else benchmark_config.num_examples),
limit=(-1 if benchmark_config.num_examples is None else benchmark_config.num_examples),
)
res = await self.evaluate_rows(
benchmark_id=benchmark_id,
input_rows=all_rows.rows,
input_rows=all_rows.data,
scoring_functions=scoring_functions,
benchmark_config=benchmark_config,
)

6
llama_stack/providers/inline/post_training/torchtune/recipes/lora_finetuning_single_device.py
View file

@ -328,13 +328,13 @@ class LoraFinetuningSingleDevice:
batch_size: int,
) -> Tuple[DistributedSampler, DataLoader]:
async def fetch_rows(dataset_id: str):
return await self.datasetio_api.get_rows_paginated(
return await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
all_rows = await fetch_rows(dataset_id)
rows = all_rows.rows
rows = all_rows.data
await validate_input_dataset_schema(
datasets_api=self.datasets_api,

10
llama_stack/providers/inline/scoring/basic/scoring.py
View file

@ -24,7 +24,9 @@ from llama_stack.providers.utils.common.data_schema_validator import (
from .config import BasicScoringConfig
from .scoring_fn.bfcl_scoring_fn import BFCLScoringFn
from .scoring_fn.equality_scoring_fn import EqualityScoringFn
from .scoring_fn.regex_parser_math_response_scoring_fn import RegexParserMathResponseScoringFn
from .scoring_fn.regex_parser_math_response_scoring_fn import (
RegexParserMathResponseScoringFn,
)
from .scoring_fn.regex_parser_scoring_fn import RegexParserScoringFn
from .scoring_fn.subset_of_scoring_fn import SubsetOfScoringFn
@ -82,12 +84,12 @@ class BasicScoringImpl(
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.scoring.value))
all_rows = await self.datasetio_api.get_rows_paginated(
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
res = await self.score(
input_rows=all_rows.rows,
input_rows=all_rows.data,
scoring_functions=scoring_functions,
)
if save_results_dataset:

6
llama_stack/providers/inline/scoring/braintrust/braintrust.py
View file

@ -167,11 +167,11 @@ class BraintrustScoringImpl(
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.scoring.value))
all_rows = await self.datasetio_api.get_rows_paginated(
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
res = await self.score(input_rows=all_rows.rows, scoring_functions=scoring_functions)
res = await self.score(input_rows=all_rows.data, scoring_functions=scoring_functions)
if save_results_dataset:
# TODO: persist and register dataset on to server for reading
# self.datasets_api.register_dataset()

6
llama_stack/providers/inline/scoring/llm_as_judge/scoring.py
View file

@ -72,12 +72,12 @@ class LlmAsJudgeScoringImpl(
dataset_def = await self.datasets_api.get_dataset(dataset_id=dataset_id)
validate_dataset_schema(dataset_def.dataset_schema, get_valid_schemas(Api.scoring.value))
all_rows = await self.datasetio_api.get_rows_paginated(
all_rows = await self.datasetio_api.iterrows(
dataset_id=dataset_id,
rows_in_page=-1,
limit=-1,
)
res = await self.score(
input_rows=all_rows.rows,
input_rows=all_rows.data,
scoring_functions=scoring_functions,
)
if save_results_dataset:

65
llama_stack/providers/remote/datasetio/huggingface/huggingface.py
View file

@ -4,13 +4,13 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Optional
from urllib.parse import parse_qs, urlparse
import datasets as hf_datasets
from llama_stack.apis.datasetio import DatasetIO, PaginatedRowsResult
from llama_stack.apis.datasetio import DatasetIO, IterrowsResponse
from llama_stack.apis.datasets import Dataset
from llama_stack.providers.datatypes import DatasetsProtocolPrivate
from llama_stack.providers.utils.datasetio.url_utils import get_dataframe_from_url
from llama_stack.providers.utils.kvstore import kvstore_impl
from .config import HuggingfaceDatasetIOConfig
@ -18,22 +18,14 @@ from .config import HuggingfaceDatasetIOConfig
DATASETS_PREFIX = "datasets:"
def load_hf_dataset(dataset_def: Dataset):
if dataset_def.metadata.get("path", None):
dataset = hf_datasets.load_dataset(**dataset_def.metadata)
else:
df = get_dataframe_from_url(dataset_def.url)
def parse_hf_params(dataset_def: Dataset):
uri = dataset_def.source.uri
parsed_uri = urlparse(uri)
params = parse_qs(parsed_uri.query)
params = {k: v[0] for k, v in params.items()}
path = parsed_uri.path.lstrip("/")
if df is None:
raise ValueError(f"Failed to load dataset from {dataset_def.url}")
dataset = hf_datasets.Dataset.from_pandas(df)
# drop columns not specified by schema
if dataset_def.dataset_schema:
dataset = dataset.select_columns(list(dataset_def.dataset_schema.keys()))
return dataset
return path, params
class HuggingfaceDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
@ -64,7 +56,7 @@ class HuggingfaceDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
key = f"{DATASETS_PREFIX}{dataset_def.identifier}"
await self.kvstore.set(
key=key,
value=dataset_def.json(),
value=dataset_def.model_dump_json(),
)
self.dataset_infos[dataset_def.identifier] = dataset_def
@ -73,41 +65,34 @@ class HuggingfaceDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
await self.kvstore.delete(key=key)
del self.dataset_infos[dataset_id]
async def get_rows_paginated(
async def iterrows(
self,
dataset_id: str,
rows_in_page: int,
page_token: Optional[str] = None,
filter_condition: Optional[str] = None,
) -> PaginatedRowsResult:
start_index: Optional[int] = None,
limit: Optional[int] = None,
) -> IterrowsResponse:
dataset_def = self.dataset_infos[dataset_id]
loaded_dataset = load_hf_dataset(dataset_def)
path, params = parse_hf_params(dataset_def)
loaded_dataset = hf_datasets.load_dataset(path, **params)
if page_token and not page_token.isnumeric():
raise ValueError("Invalid page_token")
start_index = start_index or 0
if page_token is None or len(page_token) == 0:
next_page_token = 0
else:
next_page_token = int(page_token)
start = next_page_token
if rows_in_page == -1:
if limit is None or limit == -1:
end = len(loaded_dataset)
else:
end = min(start + rows_in_page, len(loaded_dataset))
end = min(start_index + limit, len(loaded_dataset))
rows = [loaded_dataset[i] for i in range(start, end)]
rows = [loaded_dataset[i] for i in range(start_index, end)]
return PaginatedRowsResult(
rows=rows,
total_count=len(rows),
next_page_token=str(end),
return IterrowsResponse(
data=rows,
next_start_index=end if end < len(loaded_dataset) else None,
)
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None:
dataset_def = self.dataset_infos[dataset_id]
loaded_dataset = load_hf_dataset(dataset_def)
path, params = parse_hf_params(dataset_def)
loaded_dataset = hf_datasets.load_dataset(path, **params)
# Convert rows to HF Dataset format
new_dataset = hf_datasets.Dataset.from_list(rows)

17
llama_stack/providers/utils/datasetio/url_utils.py
View file

@ -10,18 +10,17 @@ from urllib.parse import unquote
import pandas
from llama_stack.apis.common.content_types import URL
from llama_stack.providers.utils.memory.vector_store import parse_data_url
def get_dataframe_from_url(url: URL):
def get_dataframe_from_uri(uri: str):
df = None
if url.uri.endswith(".csv"):
df = pandas.read_csv(url.uri)
elif url.uri.endswith(".xlsx"):
df = pandas.read_excel(url.uri)
elif url.uri.startswith("data:"):
parts = parse_data_url(url.uri)
if uri.endswith(".csv"):
df = pandas.read_csv(uri)
elif uri.endswith(".xlsx"):
df = pandas.read_excel(uri)
elif uri.startswith("data:"):
parts = parse_data_url(uri)
data = parts["data"]
if parts["is_base64"]:
data = base64.b64decode(data)
@ -39,6 +38,6 @@ def get_dataframe_from_url(url: URL):
else:
df = pandas.read_excel(data_bytes)
else:
raise ValueError(f"Unsupported file type: {url}")
raise ValueError(f"Unsupported file type: {uri}")
return df

76
llama_stack/templates/open-benchmark/open_benchmark.py
View file

@ -6,7 +6,7 @@
from typing import Dict, List, Tuple
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.datasets import DatasetPurpose, URIDataSource
from llama_stack.apis.models.models import ModelType
from llama_stack.distribution.datatypes import (
BenchmarkInput,
@ -171,76 +171,42 @@ def get_distribution_template() -> DistributionTemplate:
DatasetInput(
dataset_id="simpleqa",
provider_id="huggingface",
url=URL(uri="https://huggingface.co/datasets/llamastack/simpleqa"),
metadata={
"path": "llamastack/simpleqa",
"split": "train",
},
dataset_schema={
"input_query": {"type": "string"},
"expected_answer": {"type": "string"},
"chat_completion_input": {"type": "string"},
},
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/simpleqa?split=train",
),
),
DatasetInput(
dataset_id="mmlu_cot",
provider_id="huggingface",
url=URL(uri="https://huggingface.co/datasets/llamastack/mmlu_cot"),
metadata={
"path": "llamastack/mmlu_cot",
"name": "all",
"split": "test",
},
dataset_schema={
"input_query": {"type": "string"},
"expected_answer": {"type": "string"},
"chat_completion_input": {"type": "string"},
},
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/mmlu_cot?split=test&name=all",
),
),
DatasetInput(
dataset_id="gpqa_cot",
provider_id="huggingface",
url=URL(uri="https://huggingface.co/datasets/llamastack/gpqa_0shot_cot"),
metadata={
"path": "llamastack/gpqa_0shot_cot",
"name": "gpqa_main",
"split": "train",
},
dataset_schema={
"input_query": {"type": "string"},
"expected_answer": {"type": "string"},
"chat_completion_input": {"type": "string"},
},
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/gpqa_0shot_cot?split=test&name=gpqa_main",
),
),
DatasetInput(
dataset_id="math_500",
provider_id="huggingface",
url=URL(uri="https://huggingface.co/datasets/llamastack/math_500"),
metadata={
"path": "llamastack/math_500",
"split": "test",
},
dataset_schema={
"input_query": {"type": "string"},
"expected_answer": {"type": "string"},
"chat_completion_input": {"type": "string"},
},
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/math_500?split=test",
),
),
DatasetInput(
dataset_id="bfcl",
provider_id="huggingface",
url=URL(uri="https://huggingface.co/datasets/llamastack/bfcl_v3"),
metadata={
"path": "llamastack/bfcl_v3",
"split": "train",
},
dataset_schema={
"function": {"type": "string"},
"language": {"type": "string"},
"ground_truth": {"type": "string"},
"id": {"type": "string"},
"chat_completion_input": {"type": "string"},
},
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/bfcl_v3?split=train",
),
),
]

91
llama_stack/templates/open-benchmark/run.yaml
View file

@ -158,80 +158,39 @@ shields:
- shield_id: meta-llama/Llama-Guard-3-8B
vector_dbs: []
datasets:
- dataset_schema:
input_query:
type: string
expected_answer:
type: string
chat_completion_input:
type: string
url:
uri: https://huggingface.co/datasets/llamastack/simpleqa
metadata:
path: llamastack/simpleqa
split: train
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/simpleqa?split=train
metadata: {}
dataset_id: simpleqa
provider_id: huggingface
- dataset_schema:
input_query:
type: string
expected_answer:
type: string
chat_completion_input:
type: string
url:
uri: https://huggingface.co/datasets/llamastack/mmlu_cot
metadata:
path: llamastack/mmlu_cot
name: all
split: test
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/mmlu_cot?split=test&name=all
metadata: {}
dataset_id: mmlu_cot
provider_id: huggingface
- dataset_schema:
input_query:
type: string
expected_answer:
type: string
chat_completion_input:
type: string
url:
uri: https://huggingface.co/datasets/llamastack/gpqa_0shot_cot
metadata:
path: llamastack/gpqa_0shot_cot
name: gpqa_main
split: train
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/gpqa_0shot_cot?split=test&name=gpqa_main
metadata: {}
dataset_id: gpqa_cot
provider_id: huggingface
- dataset_schema:
input_query:
type: string
expected_answer:
type: string
chat_completion_input:
type: string
url:
uri: https://huggingface.co/datasets/llamastack/math_500
metadata:
path: llamastack/math_500
split: test
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/math_500?split=test
metadata: {}
dataset_id: math_500
provider_id: huggingface
- dataset_schema:
function:
type: string
language:
type: string
ground_truth:
type: string
id:
type: string
chat_completion_input:
type: string
url:
uri: https://huggingface.co/datasets/llamastack/bfcl_v3
metadata:
path: llamastack/bfcl_v3
split: train
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/bfcl_v3?split=train
metadata: {}
dataset_id: bfcl
provider_id: huggingface
scoring_fns: []

3
llama_stack/templates/template.py
View file

@ -11,6 +11,7 @@ import jinja2
import yaml
from pydantic import BaseModel, Field
from llama_stack.apis.datasets import DatasetPurpose
from llama_stack.apis.models.models import ModelType
from llama_stack.distribution.datatypes import (
Api,
@ -214,7 +215,9 @@ class DistributionTemplate(BaseModel):
# Register YAML representer for ModelType
yaml.add_representer(ModelType, enum_representer)
yaml.add_representer(DatasetPurpose, enum_representer)
yaml.SafeDumper.add_representer(ModelType, enum_representer)
yaml.SafeDumper.add_representer(DatasetPurpose, enum_representer)
for output_dir in [yaml_output_dir, doc_output_dir]:
output_dir.mkdir(parents=True, exist_ok=True)