Hugging face dataset streaming support (#177)

* streaming dataset
* no stride / offset for streaming
* add recipe example for streaming

Author: vince62s

eole/config/data.py

@@ -210,7 +210,7 @@ def _validate_vocab_config(self, build_vocab_only=False):
     @staticmethod
     def _validate_file(file_path, info):
         """Check `file_path` is valid or raise `IOError`."""
-        if file_path == "dummy":
+        if file_path == "dummy" or file_path.startswith("hf://"):
             # hack to allow creating objects with required fields
             pass
         elif not os.path.isfile(file_path):

eole/inputters/text_corpus.py

@@ -1,12 +1,14 @@
 """Module that contain shard utils for dynamic data."""
 
 import os
+import re
 from eole.utils.logging import logger
 from eole.constants import CorpusName, CorpusTask
 from eole.transforms import TransformPipe
 from eole.inputters.text_utils import transform_bucket
 from contextlib import contextmanager
 import itertools
+from datasets import load_dataset
 
 
 @contextmanager
@@ -102,11 +104,38 @@ def __init__(self, name, src, tgt, sco=None, align=None):
         self.sco = sco
         self.align = align
 
+    def _is_hf_dataset(self, path):
+        """
+        Check if a given path refers to a Hugging Face dataset.
+        Matchs the 'hf://' prefix and assumes the dataset is in streaming mode.
+        Match the last '/field' to get the language / score field
+        """
+        pattern = r"hf://([^/]+/[^/]+)/([^/]+)"
+        if isinstance(path, str):
+            return re.match(pattern, path)
+        else:
+            return None
+
+    def _load_hf_dataset(self, path):
+        """
+        Load a Hugging Face dataset from the given identifier.
+        Matchs the 'hf://' prefix and assumes the dataset is in streaming mode.
+        Match the last '/field' to get the language / score field
+        """
+        pattern = r"hf://([^/]+/[^/]+)/([^/]+)"
+        dataset_name = re.match(pattern, self.src).group(1)
+        return load_dataset(dataset_name, split="train", streaming=True)
+
     def load(self, offset=0, stride=1):
         """
         Load file and iterate by lines.
         `offset` and `stride` allow to iterate only on every
         `stride` example, starting from `offset`.
+        In the case of local files, all files are open exactly the same way by each worker
+        Therefore we need to apply a stride / offset rule to make sure we do not process the same ex.
+        In the case of HF streaming mode we need to make sure we have more shards than workers.
+        Typically we recommend to have shard being a multiple of workers for instance for big datasets:
+        16 shards for 4 workers. The shards will be iterated automatically since HF locks shards when in use.
         """
 
         def make_ex(sline, tline, scoline, align):
@@ -133,6 +162,16 @@ def make_ex(sline, tline, scoline, align):
                     if scoline is None:
                         scoline = 1.0
                     yield make_ex(sline, tline, scoline, align)
+
+        elif self._is_hf_dataset(self.src):
+            # If `src` is a Hugging Face dataset identifier
+            dataset = self._load_hf_dataset(self.src)
+            for i, example in enumerate(dataset):
+                sline = example.get(self.src.split("/")[-1])
+                tline = example.get(self.tgt.split("/")[-1])
+                scoline = example.get(self.sco.split("/")[-1], 1.0)
+                yield make_ex(sline, tline, scoline, None)
+
         else:
             with exfile_open(self.src, mode="rb") as fs, exfile_open(self.tgt, mode="rb") as ft, exfile_open(
                 self.sco, mode="rb"

recipes/NewsPalm-synthetic/README.md

@@ -0,0 +1,42 @@
+# Example of using Hugging Face streaming dataset
+
+## Based on: https://arxiv.org/pdf/2408.06537
+
+Introducing the NewsPaLM MBR and QE Dataset:
+LLM-Generated High-Quality Parallel Data Outperforms Traditional
+Web-Crawled Data
+
+### Get the vocab and BPE model on HF
+
+https://huggingface.co/eole-nlp/NewsPalmSynthetic-ENDE
+
+copy files: 
+* ende.vocab2
+* subwords.en_de.bpe
+
+
+### Optionally you can get the trained model to test it.
+
+
+* config.json
+* vocab.json
+* model.00.safetensors
+
+## Train with the yaml config file
+
+```
+eole train -c newspalm-synthetic-hfstreaming.yaml
+```
+
+## Start the gradio based translator
+
+```
+eole predict -c inference.yaml --src newstest2023-src.en --output newstest2023-hyp.de
+```
+
+Then you can score with sacrebleu and/or comet
+
+
+Scoring with Unbabel/wmt22-comet-da gives: 81.90
+
+You can compare to table 5 lines 2a) to 2d) of the paper https://arxiv.org/pdf/2408.06537

recipes/NewsPalm-synthetic/inference.yaml

@@ -0,0 +1,15 @@
+# Model info
+model_path: "Path to model.00.safetensors"
+# Inference
+max_length: 1024
+max_length_ratio: 3
+world_size: 1
+gpu_ranks: [0]
+batch_type: tokens
+batch_size: 16384
+compute_dtype: fp16
+beam_size: 4
+n_best: 1
+report_time: true
+self_attn_backend: "pytorch"
+src: none

recipes/NewsPalm-synthetic/newspalm-synthetic-hfstreaming.yaml

@@ -0,0 +1,132 @@
+seed: 1234
+share_vocab: true
+src_vocab: "ende.vocab2"
+
+src_words_min_frequency: 1
+vocab_size_multiple: 8
+report_every: 100
+skip_empty_level: silent
+valid_metrics: ["BLEU"]
+scoring_debug: True
+
+# transforms config
+transforms: [onmt_tokenize, filtertoolong]
+transforms_configs:
+  onmt_tokenize:
+    #### Subword
+    src_subword_type: bpe
+    src_subword_model: "subwords.en_de.bpe"
+    src_onmttok_kwargs: {"mode": "aggressive", "joiner_annotate": True, "preserve_placeholders": True, "case_markup": True, "soft_case_regions": True, "preserve_segmented_tokens": True, "segment_case": True, "segment_numbers": True, "segment_alphabet_change": True}
+    tgt_subword_type: bpe
+    tgt_subword_model: "subwords.en_de.bpe"
+    tgt_onmttok_kwargs: {"mode": "aggressive", "joiner_annotate": True, "preserve_placeholders": True, "case_markup": True, "soft_case_regions": True, "preserve_segmented_tokens": True, "segment_case": True, "segment_numbers": True, "segment_alphabet_change": True}
+
+  filtertoolong:
+    src_seq_length: 1024
+    tgt_seq_length: 1024
+
+# Corpus opts:
+data:
+    synth-mbr-decoded-sentlevel:
+# 997834 ex - 315MB
+        path_src: "hf://eole-nlp/synth-mbr-decoded-sentlevel/en"
+        path_tgt: "hf://eole-nlp/synth-mbr-decoded-sentlevel/de"
+        path_sco: "hf://eole-nlp/synth-mbr-decoded-sentlevel/sco"
+        transforms: [onmt_tokenize, filtertoolong]
+        weight: 12
+
+    synth-greedy-decoded-sentlevel:
+# 832709 ex - 250MB
+        path_src: "hf://eole-nlp/synth-greedy-decoded-sentlevel/en"
+        path_tgt: "hf://eole-nlp/synth-greedy-decoded-sentlevel/de"
+        path_sco: "hf://eole-nlp/synth-greedy-decoded-sentlevel/sco"
+        transforms: [onmt_tokenize, filtertoolong]
+        weight: 10
+
+    synth-qe-reranked-doclevel:
+# 417102 ex - 970MB
+        path_src: "hf://eole-nlp/synth-qe-reranked-doclevel/en"
+        path_tgt: "hf://eole-nlp/synth-qe-reranked-doclevel/de"
+        path_sco: "hf://eole-nlp/synth-qe-reranked-doclevel/sco"
+        transforms: [onmt_tokenize, filtertoolong]
+        weight: 1
+
+    synth-greedy-decoded-doclevel:
+# 857937 ex - 1.7GB
+        path_src: "hf://eole-nlp/europarl-v10.de-en/en"
+        path_tgt: "hf://eole-nlp/europarl-v10.de-en/de"
+        path_sco: "hf://eole-nlp/europarl-v10.de-en/sco"
+        transforms: [onmt_tokenize, filtertoolong]
+        weight: 2
+
+    valid:
+        path_src: "newstest2023-src.en"
+        path_tgt: "newstest2023-ref.de"
+        transforms: [onmt_tokenize]
+
+training:
+    # General opts
+    torch_compile: false
+    
+    model_path: "6-6-16-1024-4096-hfstreaming"
+    keep_checkpoint: 50
+    save_checkpoint_steps: 5000
+    average_decay: 0.0005
+    train_steps: 51000
+    valid_steps: 100
+
+    # Batching
+    bucket_size: 10000
+    num_workers: 4
+    prefetch_factor: 400
+    world_size: 1
+    gpu_ranks: [0]
+    batch_type: "tokens"
+    batch_size: 12144
+    valid_batch_size: 8192
+    batch_size_multiple: 1
+    accum_count: [6, 6, 6]
+    accum_steps: [0, 15000, 30000]
+
+    # Optimization
+    compute_dtype: "fp16"
+    apex_opt_level: ""
+    optim: "adamw"
+    reset_optim: "all"
+    learning_rate: 1
+    warmup_steps: 6000
+    decay_method: "noam"
+    adam_beta2: 0.998
+    max_grad_norm: 1
+    label_smoothing: 0.1
+    param_init_method: "xavier_uniform"
+    normalization: "tokens"
+
+    dropout_steps: [0, 15000, 30000]
+    dropout: [0.1, 0.1, 0.1]
+    attention_dropout: [0.0, 0.0, 0.0]
+    score_threshold: 0.65
+
+    freeze_decoder: false
+    freeze_encoder: false
+
+model:
+    architecture: "transformer"
+    layers: 6
+    hidden_size: 1024
+    heads: 16
+    transformer_ff: 4096
+    add_qkvbias: false
+    add_ffnbias: true
+    mlp_activation_fn: gated-silu
+    add_estimator: false
+    share_decoder_embeddings: true
+    share_embeddings: true
+    layer_norm: standard
+    norm_eps: 1e-6
+    rope_config:
+        rotary_interleave: false
+    embeddings:
+        word_vec_size: 1024
+        position_encoding_type: "Rotary"
+        freeze_word_vecs_dec: false