Add preliminary Dolma v1.7 configurations, fix corner case in tokens. (…

…#120) * data * added configs * adding experiments for blocklist * fixed bug in tokenizer * removed models dir for now
allenai · Feb 13, 2024 · c6c0b47 · c6c0b47
1 parent b9a8b94
commit c6c0b47
Show file tree

Hide file tree

Showing 11 changed files with 510 additions and 19 deletions.
diff --git a/configs/dolma-v1_7/v1_5-baseline/300g_sample.yaml b/configs/dolma-v1_7/v1_5-baseline/300g_sample.yaml
@@ -0,0 +1,97 @@
+
+streams:
+  - name: books
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/books/*.gz
+    attributes: &attributes
+      - paloma_paragraphs
+      - paloma_documents
+      - random_number_v1
+    output: &output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/books
+      max_size_in_bytes: 4294967296
+      discard_fields:
+        - attributes
+    filter: &filter
+      include:
+          - "$.attributes[?(@.random_number_v1__random_number_v1__random[0][2] < 0.10)]"
+      exclude:
+        - "[email protected][?(@.paloma_documents_bff_duplicates && @.paloma_documents_bff_duplicates[0] && @.paloma_documents_bff_duplicates[0][2] >= 1.0)]"
+        - "[email protected][?(@.paloma_paragraphs_bff_duplicates && @.paloma_paragraphs_bff_duplicates[0] && @.paloma_paragraphs_bff_duplicates[0][2] >= 1.0)]"
+
+  - name: c4
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/c4/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/c4
+    filter: *filter
+
+  - name: cc_en_head
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_head/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/cc_en_head
+    filter: *filter
+
+  - name: cc_en_middle
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_middle/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/cc_en_middle
+    filter: *filter
+
+  - name: cc_en_tail
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_tail/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/cc_en_tail
+    filter: *filter
+
+  - name: pes2o
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/pes2o/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/pes2o
+    filter: *filter
+
+  - name: reddit
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/reddit/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/reddit
+    filter: *filter
+
+  - name: stack
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/stack/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/stack
+    filter: *filter
+
+  - name: wiki
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/wiki/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/wiki
+    filter: *filter
+
+work_dir:
+  input: "/tmp/olmo-mix-v1_5/input"
+  output: "/tmp/olmo-mix-v1_5/output"
+processes: 188
diff --git a/configs/dolma-v1_7/v1_5-baseline/300g_tok.yml b/configs/dolma-v1_7/v1_5-baseline/300g_tok.yml
@@ -0,0 +1,23 @@
+destination: ${oc.env:HOME}/ai2-llm/preprocessed/olmo-mix/v1_5-300G-decon/gpt-neox-olmo-dolma-v1_6
+documents:
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/books
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/c4
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/cc_en_head
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/cc_en_middle
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/cc_en_tail
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/pes2o
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/reddit
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/stack
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5-300G-decon/documents/wiki
+  # - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/starcoder/v0/documents/*/*.json.gz
+
+processes: 40
+seed: 3920
+max_size: 21_474_836_480
+
+tokenizer:
+  name_or_path: allenai/gpt-neox-olmo-dolma-v1_5
+  bos_token_id: null
+  eos_token_id: 50279
+  pad_token_id: 1
+  segment_before_tokenization: false
diff --git a/configs/dolma-v1_7/v1_5-baseline/step2-run-decontamination/dolma-v1_5_docs.yaml b/configs/dolma-v1_7/v1_5-baseline/step2-run-decontamination/dolma-v1_5_docs.yaml
@@ -0,0 +1,26 @@
+documents:
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/books/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/c4/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_head/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_middle/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_tail/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/pes2o/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/reddit/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/stack/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/wiki/*.gz
+
+
+dedupe:
+  name: paloma_documents
+  documents:
+    attribute_name: paloma_documents_bff_duplicates
+    key: $.text
+  skip_empty: true
+
+bloom_filter:
+  read_only: true
+  estimated_doc_count: 188815
+  desired_false_positive_rate: 1e-15
+  file: ${oc.env:HOME}/perplexity/filters/paloma_documents.bin
+
+processes: 94
diff --git a/configs/dolma-v1_7/v1_5-baseline/step2-run-decontamination/dolma-v1_5_para.yaml b/configs/dolma-v1_7/v1_5-baseline/step2-run-decontamination/dolma-v1_5_para.yaml
@@ -0,0 +1,29 @@
+documents:
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/books/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/c4/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_head/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_middle/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/cc_en_tail/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/pes2o/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/reddit/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/stack/*.gz
+  - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5/documents/wiki/*.gz
+  # - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/books/*.gz
+  # - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/wiki/*.gz
+
+
+dedupe:
+  name: paloma_paragraphs
+  paragraphs:
+    attribute_name: paloma_paragraphs_bff_duplicates
+  skip_empty: true
+
+bloom_filter:
+  read_only: true
+  estimated_doc_count: 2336120
+  # size_in_bytes: 104857600  # 100 MB; smaller causes too many FPs
+  desired_false_positive_rate: 1e-15
+  # file: s3://ai2-llm/bloom-filters/perplexity-suite-v3_option2.bin
+  file: ${oc.env:HOME}/perplexity/filters/paloma_paragraphs.bin
+
+processes: 94
diff --git a/configs/dolma-v1_7/v1_6-baseline/300g_sample.yml b/configs/dolma-v1_7/v1_6-baseline/300g_sample.yml
@@ -0,0 +1,96 @@
+streams:
+  - name: books
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/books/*.gz
+    attributes: &attributes
+      - paloma_paragraphs
+      - paloma_documents
+      - random_number_v1
+    output: &output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/books
+      max_size_in_bytes: 38949672960
+      discard_fields:
+        - attributes
+    filter: &filter
+      include:
+        - "$.attributes[?(@.random_number_v1__random_number_v1__random[0][2] < 0.10)]"
+      exclude:
+        - "[email protected][?(@.paloma_documents_bff_duplicates && @.paloma_documents_bff_duplicates[0] && @.paloma_documents_bff_duplicates[0][2] >= 1.0)]"
+        - "[email protected][?(@.paloma_paragraphs_bff_duplicates && @.paloma_paragraphs_bff_duplicates[0] && @.paloma_paragraphs_bff_duplicates[0][2] >= 1.0)]"
+
+  - name: c4
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/c4/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/c4
+    filter: *filter
+
+  - name: cc_en_head
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/cc_en_head/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/cc_en_head
+    filter: *filter
+
+  - name: cc_en_middle
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/cc_en_middle/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/cc_en_middle
+    filter: *filter
+
+  - name: cc_en_tail
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/cc_en_tail/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/cc_en_tail
+    filter: *filter
+
+  - name: pes2o
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/pes2o/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/pes2o
+    filter: *filter
+
+  - name: reddit
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/reddit/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/reddit
+    filter: *filter
+
+  - name: stack
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/stack/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/stack
+    filter: *filter
+
+  - name: wiki
+    documents:
+      - s3://ai2-llm/pretraining-data/sources/olmo-mix/v1_5r2/documents/wiki/*.gz
+    attributes: *attributes
+    output:
+      <<: *output
+      path: ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/wiki
+    filter: *filter
+
+work_dir:
+  input: "/tmp/olmo-mix-v1_6/input"
+  output: "/tmp/olmo-mix-v1_6/output"
+processes: 188
diff --git a/configs/dolma-v1_7/v1_6-baseline/300g_tok.yml b/configs/dolma-v1_7/v1_6-baseline/300g_tok.yml
@@ -0,0 +1,23 @@
+destination: ${oc.env:HOME}/ai2-llm/preprocessed/olmo-mix/v1_6-300G-decon/gpt-neox-olmo-dolma-v1_6
+documents:
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/books
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/c4
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/cc_en_head
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/cc_en_middle
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/cc_en_tail
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/pes2o
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/reddit
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/stack
+  - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/wiki
+  # - ${oc.env:HOME}/ai2-llm/pretraining-data/sources/starcoder/v0/documents/*/*.json.gz
+
+processes: 40
+seed: 3920
+max_size: 21_474_836_480
+
+tokenizer:
+  name_or_path: allenai/gpt-neox-olmo-dolma-v1_5
+  bos_token_id: null
+  eos_token_id: 50279
+  pad_token_id: 1
+  segment_before_tokenization: false
diff --git a/configs/dolma-v1_7/v1_6-baseline/tok_per_source.sh b/configs/dolma-v1_7/v1_6-baseline/tok_per_source.sh
@@ -0,0 +1,31 @@
+#!/bin/env bash
+
+sources=(
+    "books,1"
+    "c4,5"
+    "cc_en_head,10"
+    "cc_en_middle,10"
+    "cc_en_tail,10"
+    "pes2o,3"
+    "reddit,2"
+    "stack,5"
+    "wiki,1"
+)
+
+set -x
+
+for i in "${!sources[@]}"; do
+    # split source and number of processes
+    source=$(echo "${sources[$i]}" | cut -d',' -f1)
+    processes=$(echo "${sources[$i]}" | cut -d',' -f2)
+
+    dolma tokens \
+        --destination "s3://ai2-llm/preprocessed/olmo-mix/v1_6-300G-decon/gpt-neox-olmo-dolma-v1_6_persource/${source}" \
+        --documents "${HOME}/ai2-llm/pretraining-data/sources/olmo-mix/v1_6-300G-decon/documents/${source}" \
+        --tokenizer.name_or_path "allenai/gpt-neox-olmo-dolma-v1_5" \
+        --tokenizer.eos_token_id 50279 \
+        --tokenizer.pad_token_id 1 \
+        --processes  ${processes} \
+        --seed 3920 \
+        --max_size "21_474_836_480"
+done