Add data loader for HF oasst1 (#2951)

Make it possible to work with the OASST1 dataset directly from the HuggingFace hub.
Add  a new `hf_dataset_name` parameter to the `load_oasst_export` function.

---------

Co-authored-by: grgau <[email protected]>
Co-authored-by: Tobias Pitters <[email protected]>
Co-authored-by: Andreas Köpf <[email protected]>

Author: 4 people

model/README.md

@@ -16,15 +16,49 @@ export DATA_PATH=$PWD/.cache
 export MODEL_PATH=$PWD/.saved_models
 ```
 
-2. Then download the OA data.
+2. Then download the OA message tree JSONL file or declare the HuggingFace
+   dataset to use.
+
+Create a new or modify an existing configuration section in the `config.yaml`
+(SFT), `config_rm.yaml` (RM) or `config_rl.yaml` (RL) YAML configuration files
+located in the `model_training/configs/` directory and specify the OA JSONL data
+file or HuggingFace dataset to use.
+
+- To use a local OASST JSONL file (either `.jsonl` or `.jsonl.gz`) specify the
+  file name with the `input_file_path` configuration option. Place the file
+  either in the `cache_dir` (`DATA_PATH`) or specify an absolute path.
 
 ```bash
-cp /path/to/<oa.jsonl> $DATA_PATH
+cp /path/to/<oasst.trees.jsonl> $DATA_PATH
+```
+
+Example:
+
+```yaml
+my_data_config:
+  datasets:
+    - oasst_export:
+      input_file_path: oasst_export.trees.jsonl.gz
 ```
 
-Change the `<oa.jsonl>` file used in the `model_training/configs/config.yaml`,
-`model_training/configs/config_rl.yaml` and `reward/instructor/rank_datasets.py`
-files.
+- To use a HuggingFace dataset specify the dataset name with the
+  `hf_dataset_name` configuration option.
+
+Example:
+
+```yaml
+my_data_config:
+  datasets:
+    - oasst_export:
+      hf_dataset_name: OpenAssistant/oasst1
+```
+
+_Note_: If both `hf_dataset_name` and `input_file_path` are specified
+`input_file_path` will take precedence.
+
+See the
+[OpenAssistant/oasst1](https://huggingface.co/datasets/OpenAssistant/oasst1)
+dataset card on the HuggingFace hub for more information.
 
 - (TODO) add better parsing of the config files that is consistent for sft, rm
   and rl training.

model/model_training/configs/config.yaml

@@ -186,7 +186,9 @@ oasst_only:
   datasets:
     - oasst_export:
         lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk"
-        input_file_path: 2023-04-04_oasst_ready.jsonl.gz
+        hf_dataset_name: OpenAssistant/oasst1
+        #input_file_path: 2023-04-12_oasst_ready.trees.jsonl.gz
+        #top_k: 1
         val_split: 0.05
   sort_by_length: false
   use_custom_sampler: false
@@ -206,14 +208,28 @@ oasst_export_eu:
   datasets:
     - oasst_export:
         lang: "en,es,de,fr"
-        input_file_path: 2023-03-27_oasst_research_ready_synth.jsonl.gz
+        hf_dataset_name: OpenAssistant/oasst1
+    - gpt4all
+    - alpaca
+    - code_alpaca
+    - oig_file:
+        source_url: https://huggingface.co/datasets/laion/OIG/resolve/main/unified_chip2.jsonl
+        max_count: 10000
+        min_length: 100
+        val_split: 0.1
+    - oig_file:
+        source_url: https://huggingface.co/datasets/laion/OIG/raw/main/unified_grade_school_math_instructions.jsonl
+        val_split: 0.1
+        min_length: 100
+  sort_by_length: false
+  use_custom_sampler: false
 
 oasst_export_latin_cyrillic:
   save_strategy: epoch
   datasets:
     - oasst_export:
         lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk"
-        input_file_path: 2023-03-27_oasst_research_ready_synth.jsonl.gz
+        hf_dataset_name: OpenAssistant/oasst1
     - alpaca
     - oig_file:
         source_url: https://huggingface.co/datasets/laion/OIG/resolve/main/unified_chip2.jsonl
@@ -364,7 +380,7 @@ llama-30b-sft-6:
   datasets:
     - oasst_export:
         lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk"
-        input_file_path: 2023-04-12_oasst_release_ready_synth.jsonl.gz
+        hf_dataset_name: OpenAssistant/oasst1
         val_split: 0.05
     - vicuna:
         val_split: 0.05
@@ -712,6 +728,7 @@ galactica-125m:
   gradient_accumulation_steps: 2
   per_device_train_batch_size: 4
   per_device_eval_batch_size: 4
+  dtype: fp32
 
 gpt-jt:
   learning_rate: 8e-6
@@ -761,3 +778,4 @@ debug:
   log_wandb: false
   verbose: true
   num_train_epochs: 0.2
+  dtype: fp32

model/model_training/configs/config_rm.yaml

@@ -49,7 +49,6 @@ oasst-rm-1-pythia-6.9b:
   pooling: last
   datasets:
     - augment_oasst:
-        #input_file_path: augmented_latin_cyrillic_oasst_2023-03-27.jsonl
         input_file_path: augmented_latin_cyrillic_oasst_2023-03-27_v2.jsonl
     - anthropic_rlhf:
         fraction: 0.1
@@ -98,10 +97,9 @@ oasst-rm-1-pythia-2.8b:
   datasets:
     - oasst_export:
         lang: "en,es,de,fr"
-        input_file_path: 2023-03-27_oasst_research_ready_synth.jsonl.gz
+        hf_dataset_name: OpenAssistant/oasst1
         val_split: 0.1
     - augment_oasst:
-        #input_file_path: augmented_latin_cyrillic_oasst_2023-03-27.jsonl
         input_file_path: augmented_latin_cyrillic_oasst_2023-03-27_v2.jsonl
     - anthropic_rlhf:
         fraction: 0.1
@@ -142,7 +140,7 @@ oasst-rm-1-pythia-1.4b:
   datasets:
     - oasst_export:
         lang: "en,es,de,fr"
-        input_file_path: 2023-03-27_oasst_research_ready_synth.jsonl.gz
+        hf_dataset_name: OpenAssistant/oasst1
         val_split: 0.1
     - augment_oasst:
         input_file_path: augmented_latin_cyrillic_oasst_2023-03-27.jsonl

model/model_training/custom_datasets/oasst_dataset.py

@@ -1,8 +1,9 @@
 from pathlib import Path
-from typing import Literal, Optional
+from typing import Iterable, Literal, Optional
 
 from model_training.custom_datasets.formatting import DatasetEntrySft, Role, Utterance
-from oasst_data import ExportMessageNode, read_message_trees, visit_threads_depth_first
+from oasst_data import ExportMessageNode, read_dataset_message_trees, read_message_trees, visit_threads_depth_first
+from oasst_data.schemas import ExportMessageTree
 from torch import Generator
 from torch.utils.data import Dataset, random_split
 
@@ -20,7 +21,8 @@ def __getitem__(self, index):
 
 
 def load_oasst_export(
-    input_file_path: str | Path,
+    input_file_path: Optional[str | Path] = None,
+    hf_dataset_name: Optional[str] = "OpenAssistant/oasst1",
     val_split: float = 0.2,
     lang: str = "en",
     top_k: Optional[int] = None,
@@ -31,20 +33,27 @@ def load_oasst_export(
     if mode not in ("sft", "rm", "rl"):
         raise ValueError(f"Unknown dataset mode: {mode}")
 
-    lang_codes = lang.split(",")
+    lang_codes: list[str] = lang.split(",")
 
     generator = Generator()
     generator.manual_seed(manual_seed)
 
-    if not isinstance(input_file_path, Path):
-        input_file_path = Path(input_file_path)
-    if not input_file_path.is_absolute() and data_path:
-        if not isinstance(data_path, Path):
-            data_path = Path(data_path)
-        input_file_path = data_path / input_file_path
+    tree_iter: Iterable[ExportMessageTree] = None
+    if input_file_path:
+        if not isinstance(input_file_path, Path):
+            input_file_path = Path(input_file_path)
+        if not input_file_path.is_absolute() and data_path:
+            if not isinstance(data_path, Path):
+                data_path = Path(data_path)
+            input_file_path = data_path / input_file_path
+        tree_iter = read_message_trees(input_file_path)
+    elif hf_dataset_name:
+        tree_iter = read_dataset_message_trees(hf_dataset_name, split="train+validation")
+    else:
+        raise RuntimeError("Either `input_file_path` or `hf_dataset_name` must be specified.")
 
     threads_per_tree = []
-    for tree in read_message_trees(input_file_path):
+    for tree in tree_iter:
         if tree.tree_state != "ready_for_export" or not tree.prompt.review_result or tree.prompt.lang not in lang_codes:
             continue
 
@@ -145,6 +154,9 @@ def flatten(ds: ListDataset) -> ListDataset:
     train = flatten(splits[0])
     val = flatten(splits[1])
 
-    print(f"OASST data {str(input_file_path)}: {len(train)=}, {len(val)=}")
+    if input_file_path:
+        print(f"OASST JSONL file {str(input_file_path)}: {len(train)=}, {len(val)=}")
+    else:
+        print(f"OASST HF dataset {hf_dataset_name}: {len(train)=}, {len(val)=}")
 
     return train, val

oasst-data/oasst_data/__init__.py

@@ -1,4 +1,11 @@
-from oasst_data.reader import read_message_list, read_message_tree_list, read_message_trees, read_messages
+from oasst_data.reader import (
+    read_dataset_message_trees,
+    read_dataset_messages,
+    read_message_list,
+    read_message_tree_list,
+    read_message_trees,
+    read_messages,
+)
 from oasst_data.schemas import (
     ExportMessageEvent,
     ExportMessageEventEmoji,
@@ -33,4 +40,6 @@
     "visit_messages_depth_first",
     "write_message_trees",
     "write_messages",
+    "read_dataset_message_trees",
+    "read_dataset_messages",
 ]

oasst-data/oasst_data/reader.py

@@ -4,6 +4,7 @@
 from typing import Callable, Iterable, Optional, TextIO
 
 import pydantic
+from datasets import load_dataset
 
 from .schemas import ExportMessageNode, ExportMessageTree
 
@@ -17,22 +18,24 @@ def open_jsonl_read(input_file_path: str | Path) -> TextIO:
         return input_file_path.open("r", encoding="UTF-8")
 
 
-def read_oasst_obj(line: str) -> ExportMessageTree | ExportMessageNode:
-    dict_tree = json.loads(line)
+def read_oasst_obj(obj_dict: dict) -> ExportMessageTree | ExportMessageNode:
     # validate data
-    if "message_id" in dict_tree:
-        return pydantic.parse_obj_as(ExportMessageNode, dict_tree)
-    elif "message_tree_id" in dict_tree:
-        return pydantic.parse_obj_as(ExportMessageTree, dict_tree)
+    if "message_id" in obj_dict:
+        return pydantic.parse_obj_as(ExportMessageNode, obj_dict)
+    elif "message_tree_id" in obj_dict:
+        return pydantic.parse_obj_as(ExportMessageTree, obj_dict)
 
     raise RuntimeError("Unknown object in jsonl file")
 
 
-def read_oasst_jsonl(input_file_path: str | Path) -> Iterable[ExportMessageTree | ExportMessageNode]:
+def read_oasst_jsonl(
+    input_file_path: str | Path,
+) -> Iterable[ExportMessageTree | ExportMessageNode]:
     with open_jsonl_read(input_file_path) as file_in:
         # read one object per line
         for line in file_in:
-            yield read_oasst_obj(line)
+            dict_tree = json.loads(line)
+            yield read_oasst_obj(dict_tree)
 
 
 def read_message_trees(input_file_path: str | Path) -> Iterable[ExportMessageTree]:
@@ -42,18 +45,85 @@ def read_message_trees(input_file_path: str | Path) -> Iterable[ExportMessageTre
 
 
 def read_message_tree_list(
-    input_file_path: str | Path, filter: Optional[Callable[[ExportMessageTree], bool]] = None
+    input_file_path: str | Path,
+    filter: Optional[Callable[[ExportMessageTree], bool]] = None,
 ) -> list[ExportMessageTree]:
     return [t for t in read_message_trees(input_file_path) if not filter or filter(t)]
 
 
+def convert_hf_message(row: dict) -> None:
+    emojis = row.get("emojis")
+    if emojis:
+        row["emojis"] = dict(zip(emojis["name"], emojis["count"]))
+    labels = row.get("labels")
+    if labels:
+        row["labels"] = {
+            name: {"value": value, "count": count}
+            for name, value, count in zip(labels["name"], labels["value"], labels["count"])
+        }
+
+
 def read_messages(input_file_path: str | Path) -> Iterable[ExportMessageNode]:
     for x in read_oasst_jsonl(input_file_path):
         assert isinstance(x, ExportMessageNode)
         yield x
 
 
 def read_message_list(
-    input_file_path: str | Path, filter: Optional[Callable[[ExportMessageNode], bool]] = None
+    input_file_path: str | Path,
+    filter: Optional[Callable[[ExportMessageNode], bool]] = None,
 ) -> list[ExportMessageNode]:
     return [t for t in read_messages(input_file_path) if not filter or filter(t)]
+
+
+def read_dataset_message_trees(
+    hf_dataset_name: str = "OpenAssistant/oasst1",
+    split: str = "train+validation",
+) -> Iterable[ExportMessageTree]:
+    dataset = load_dataset(hf_dataset_name, split=split)
+
+    tree_dict: dict = None
+    parents: list = None
+    for row in dataset:
+        convert_hf_message(row)
+        if row["parent_id"] is None:
+            if tree_dict:
+                tree = read_oasst_obj(tree_dict)
+                assert isinstance(tree, ExportMessageTree)
+                yield tree
+
+            tree_dict = {
+                "message_tree_id": row["message_id"],
+                "tree_state": row["tree_state"],
+                "prompt": row,
+            }
+            parents = []
+        else:
+            while parents[-1]["message_id"] != row["parent_id"]:
+                parents.pop()
+            parent = parents[-1]
+            if "replies" not in parent:
+                parent["replies"] = []
+            parent["replies"].append(row)
+
+        row.pop("message_tree_id", None)
+        row.pop("tree_state", None)
+        parents.append(row)
+
+    if tree_dict:
+        tree = read_oasst_obj(tree_dict)
+        assert isinstance(tree, ExportMessageTree)
+        yield tree
+
+
+def read_dataset_messages(
+    hf_dataset_name: str = "OpenAssistant/oasst1",
+    split: str = "train+validation",
+) -> Iterable[ExportMessageNode]:
+    dataset = load_dataset(hf_dataset_name, split=split)
+
+    for row in dataset:
+        convert_hf_message(row)
+        message = read_oasst_obj(row)
+        assert isinstance(message, ExportMessageNode)
+        yield message

oasst-data/pyproject.toml

@@ -7,7 +7,8 @@ authors = [
 ]
 dependencies = [
     "pydantic>=1.10.4",
-    "loguru==0.6.0"
+    "loguru==0.6.0",
+    "datasets>=2.12.0"
 ]
 
 [project.optional-dependencies]