added CPT code

EleutherAI · Apr 14, 2024 · 6ff3ae6 · 6ff3ae6
1 parent 5cdff76
commit 6ff3ae6
Show file tree

Hide file tree

Showing 102 changed files with 2,244 additions and 12 deletions.
diff --git a/configs/datasets/train/pile+slim_pajama_300B_each.yml b/configs/datasets/train/pile+slim_pajama_300B_each.yml
@@ -0,0 +1,25 @@
+{
+  # This will sample with equal likelihood Pile and SlimPajama:
+  "train-data-paths": [
+    "data/pile/train/pile_train",
+    'data/slim_pajama/train_300B/ArXiv/ArXiv',
+    'data/slim_pajama/train_300B/Book/Book',
+    'data/slim_pajama/train_300B/C4/C4',
+    'data/slim_pajama/train_300B/Wikipedia/Wikipedia',
+    'data/slim_pajama/train_300B/Github/Github',
+    'data/slim_pajama/train_300B/StackExchange/StackExchange',
+    'data/slim_pajama/train_300B/CommonCrawl/CommonCrawl',],
+  "train-data-weights": [
+    50.0,
+    2.2140923205,
+    2.101565663,
+    13.344249736,
+    1.9986465625,
+    2.612070528,
+    1.6855393625,
+    26.0438358255
+  ],
+  "train-dataset-name": 'pile+slim_pajama_300B_each',
+  "train-iters": 264732,
+  "lr-decay-iters": 264732,
+}
diff --git a/configs/datasets/train/pile_shard0.yml b/configs/datasets/train/pile_shard0.yml
@@ -0,0 +1,33 @@
+{ 
+  "train-data-paths": [
+    "data/pile/shard_0/shard_0_text_document",
+  ],
+  "train-data-weights": [
+    1.,
+  ],
+  "train-dataset-name": 'pile_shard0',
+  "train-iters": 1000,
+  "lr-decay-iters": 1000,
+  "is_replay_enabled": true,
+  "replay_config": {
+    "enabled": true,
+    # Have to specify idx filenames from original pretraining on tasks, as they contain the num iterations
+    # and seen indices assuming we're using the same (non-replay) seed as during pretraining
+    "replay_idx_paths_prefixes": [
+      "data/pile/shard_0/shard_0_text_document_train_0_indexmap_32160ns_2048sl_1234s",
+    ],
+    "replay_data_weights":[
+      1.00,
+    ],
+    "replay_idx_offsets": [
+      1,
+    ],
+    # Fraction of samples coming from the replay buffer, between 0 and 1.
+    "replay_fraction": 0.5,
+    # Seed and reshuffle go hand in hand. They control whether you want to see the replay data in the same order
+    # as you've seen it (done by setting reshuffle to false), and if you decide to reshuffle, what seed you should
+    # use to reshuffle the seen data.
+    "replay_seed": 1234,
+    "replay_reshuffle_idx": false,
+  },
+}
diff --git a/configs/datasets/train/pile_train.yml b/configs/datasets/train/pile_train.yml
@@ -0,0 +1,11 @@
+{ 
+  "train-data-paths": [
+    "data/pile/train/pile_train",
+  ],
+  "train-data-weights": [
+    1.,
+  ],
+  "train-dataset-name": 'pile_train',
+  "train-iters": 132366,
+  "lr-decay-iters": 132366,
+}
diff --git a/configs/datasets/train/rp.yml b/configs/datasets/train/rp.yml
@@ -0,0 +1,32 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/arxiv/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/book/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/c4/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/wikipedia/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/github/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/stackexchange/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/common_crawl/2019-30/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/common_crawl/2020-05/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/common_crawl/2021-04/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/common_crawl/2022-05/folder_train/tokenized_text_document",
+    "/gpfs/alpine/csc499/proj-shared/incite_datasets/SlimPajama/tokenized300B/train_splits/common_crawl/2023-06/folder_train/tokenized_text_document",
+  ],
+  "train-data-weights": [
+    2.5,
+    4.5,
+    15.0,
+    4.5,
+    4.5,
+    2.0,
+    13.4,
+    13.4,
+    13.4,
+    13.4,
+    13.4
+  ],
+  "train-dataset-name": 'rp',
+
+
+}
diff --git a/configs/datasets/train/slim_pajama_100B_1.yml b/configs/datasets/train/slim_pajama_100B_1.yml
@@ -0,0 +1,24 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/tokenized_train_0-100B/ArXiv/ArXiv',
+    'data/slim_pajama/tokenized_train_0-100B/Book/Book',
+    'data/slim_pajama/tokenized_train_0-100B/C4/C4',
+    'data/slim_pajama/tokenized_train_0-100B/Wikipedia/Wikipedia',
+    'data/slim_pajama/tokenized_train_0-100B/Github/Github',
+    'data/slim_pajama/tokenized_train_0-100B/StackExchange/StackExchange',
+    'data/slim_pajama/tokenized_train_0-100B/CommonCrawl/CommonCrawl',
+    ],
+  "train-data-weights": [
+    3.4703977435152775,
+    3.904381603212791,
+    25.641950653802013,
+    3.804228253591696,
+    4.9994643949282045,
+    3.1815838172641993,
+    49.99799353368582,
+  ],
+  "train-iters": 44229,
+  "lr-decay-iters": 44229,
+  "train-dataset-name": 'slim_pajama_100B_1',
+}
diff --git a/configs/datasets/train/slim_pajama_100B_1_replay5.yml b/configs/datasets/train/slim_pajama_100B_1_replay5.yml
@@ -0,0 +1,28 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/tokenized_train_0-100B/ArXiv/ArXiv',
+    'data/slim_pajama/tokenized_train_0-100B/Book/Book',
+    'data/slim_pajama/tokenized_train_0-100B/C4/C4',
+    'data/slim_pajama/tokenized_train_0-100B/Wikipedia/Wikipedia',
+    'data/slim_pajama/tokenized_train_0-100B/Github/Github',
+    'data/slim_pajama/tokenized_train_0-100B/StackExchange/StackExchange',
+    'data/slim_pajama/tokenized_train_0-100B/CommonCrawl/CommonCrawl',
+
+    'data/pile_replay_shards/replay_10B_1/splits',
+    ],
+  "train-data-weights": [
+    3.4703977435152775,
+    3.904381603212791,
+    25.641950653802013,
+    3.804228253591696,
+    4.9994643949282045,
+    3.1815838172641993,
+    49.99799353368582,
+
+    5.0
+  ],
+  "train-iters": 44229,
+  "lr-decay-iters": 44229,
+  "train-dataset-name": 'slim_pajama_100B_1_replay5',
+}
diff --git a/configs/datasets/train/slim_pajama_100B_2.yml b/configs/datasets/train/slim_pajama_100B_2.yml
@@ -0,0 +1,24 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/tokenized_train_100B-200B/ArXiv/ArXiv',
+    'data/slim_pajama/tokenized_train_100B-200B/Book/Book',
+    'data/slim_pajama/tokenized_train_100B-200B/C4/C4',
+    'data/slim_pajama/tokenized_train_100B-200B/Wikipedia/Wikipedia',
+    'data/slim_pajama/tokenized_train_100B-200B/Github/Github',
+    'data/slim_pajama/tokenized_train_100B-200B/StackExchange/StackExchange',
+    'data/slim_pajama/tokenized_train_100B-200B/CommonCrawl/CommonCrawl',
+  ],
+  "train-data-weights": [
+    4.03666599074094,
+    3.927523855378127,
+    25.467175464208918,
+    3.7984379710376293,
+    4.990226864678155,
+    3.1957646326079723,
+    49.58420522134826,
+    ],
+  "train-iters": 44229,
+  "lr-decay-iters": 44229,
+  "train-dataset-name": 'slim_pajama_100B_2',
+}
diff --git a/configs/datasets/train/slim_pajama_100B_2_replay5.yml b/configs/datasets/train/slim_pajama_100B_2_replay5.yml
@@ -0,0 +1,45 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/tokenized_train_100B-200B/ArXiv/ArXiv',
+    'data/slim_pajama/tokenized_train_100B-200B/Book/Book',
+    'data/slim_pajama/tokenized_train_100B-200B/C4/C4',
+    'data/slim_pajama/tokenized_train_100B-200B/Wikipedia/Wikipedia',
+    'data/slim_pajama/tokenized_train_100B-200B/Github/Github',
+    'data/slim_pajama/tokenized_train_100B-200B/StackExchange/StackExchange',
+    'data/slim_pajama/tokenized_train_100B-200B/CommonCrawl/CommonCrawl',
+
+    'data/pile_replay_shards/replay_10B_2/splits',
+
+    'data/sp_replay_shards/100B_1_shard1/ArXiv/ArXiv',
+    'data/sp_replay_shards/100B_1_shard1/Book/Book',
+    'data/sp_replay_shards/100B_1_shard1/C4/C4',
+    'data/sp_replay_shards/100B_1_shard1/Wikipedia/Wikipedia',
+    'data/sp_replay_shards/100B_1_shard1/Github/Github',
+    'data/sp_replay_shards/100B_1_shard1/StackExchange/StackExchange',
+    'data/sp_replay_shards/100B_1_shard1/CommonCrawl/CommonCrawl',
+  ],
+  "train-data-weights": [
+    4.03666599074094,
+    3.927523855378127,
+    25.467175464208918,
+    3.7984379710376293,
+    4.990226864678155,
+    3.1957646326079723,
+    49.58420522134826,
+
+    3.8125,
+
+    # total: 1.1875,
+    0.04337997179394097,
+    0.04880477004015989,
+    0.3205243831725252,
+    0.0475528531698962,
+    0.06249330493660256,
+    0.03976979771580249,
+    0.6249749191710727,
+    ],
+  "train-iters": 44229,
+  "lr-decay-iters": 44229,
+  "train-dataset-name": 'slim_pajama_100B_2_replay5',
+}
diff --git a/configs/datasets/train/slim_pajama_100B_3.yml b/configs/datasets/train/slim_pajama_100B_3.yml
@@ -0,0 +1,24 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/tokenized_train_200B-300B/ArXiv/ArXiv',
+    'data/slim_pajama/tokenized_train_200B-300B/Book/Book',
+    'data/slim_pajama/tokenized_train_200B-300B/C4/C4',
+    'data/slim_pajama/tokenized_train_200B-300B/Wikipedia/Wikipedia',
+    'data/slim_pajama/tokenized_train_200B-300B/Github/Github',
+    'data/slim_pajama/tokenized_train_200B-300B/StackExchange/StackExchange',
+    'data/slim_pajama/tokenized_train_200B-300B/CommonCrawl/CommonCrawl',
+  ],
+  "train-data-weights": [
+    3.491756366873565,
+    4.084283062119696,
+    25.524317038754475,
+    3.8109321899190314,
+    4.89534056131328,
+    3.254459546224121,
+    49.93891123479581,
+    ],
+  "train-iters": 44229,
+  "lr-decay-iters": 44229,
+  "train-dataset-name": 'slim_pajama_100B_3',
+}
diff --git a/configs/datasets/train/slim_pajama_100B_3_replay5.yml b/configs/datasets/train/slim_pajama_100B_3_replay5.yml
@@ -0,0 +1,61 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/tokenized_train_200B-300B/ArXiv/ArXiv',
+    'data/slim_pajama/tokenized_train_200B-300B/Book/Book',
+    'data/slim_pajama/tokenized_train_200B-300B/C4/C4',
+    'data/slim_pajama/tokenized_train_200B-300B/Wikipedia/Wikipedia',
+    'data/slim_pajama/tokenized_train_200B-300B/Github/Github',
+    'data/slim_pajama/tokenized_train_200B-300B/StackExchange/StackExchange',
+    'data/slim_pajama/tokenized_train_200B-300B/CommonCrawl/CommonCrawl',
+
+    'data/pile_replay_shards/replay_10B_3/splits',
+
+    'data/sp_replay_shards/100B_1_shard2/ArXiv/ArXiv',
+    'data/sp_replay_shards/100B_1_shard2/Book/Book',
+    'data/sp_replay_shards/100B_1_shard2/C4/C4',
+    'data/sp_replay_shards/100B_1_shard2/Wikipedia/Wikipedia',
+    'data/sp_replay_shards/100B_1_shard2/Github/Github',
+    'data/sp_replay_shards/100B_1_shard2/StackExchange/StackExchange',
+    'data/sp_replay_shards/100B_1_shard2/CommonCrawl/CommonCrawl',
+
+    'data/sp_replay_shards/100B_2_shard1/ArXiv/ArXiv',
+    'data/sp_replay_shards/100B_2_shard1/Book/Book',
+    'data/sp_replay_shards/100B_2_shard1/C4/C4',
+    'data/sp_replay_shards/100B_2_shard1/Wikipedia/Wikipedia',
+    'data/sp_replay_shards/100B_2_shard1/Github/Github',
+    'data/sp_replay_shards/100B_2_shard1/StackExchange/StackExchange',
+    'data/sp_replay_shards/100B_2_shard1/CommonCrawl/CommonCrawl',
+  ],
+  "train-data-weights": [3.491756366873565,
+    4.084283062119696,
+    25.524317038754475,
+    3.8109321899190314,
+    4.89534056131328,
+    3.254459546224121,
+    49.93891123479581,
+
+    3.088125,
+
+    # total: 0.961875,
+    0.03513777715309219,
+    0.03953186373252951,
+    0.2596247503697454,
+    0.03851781106761592,
+    0.05061957699864807,
+    0.03221353614980002,
+    0.506229684528569,
+
+    #total: 0.95,
+    0.0403666599074094,
+    0.03927523855378127,
+    0.25467175464208913,
+    0.03798437971037629,
+    0.049902268646781545,
+    0.03195764632607972,
+    0.4958420522134826,
+    ],
+  "train-iters": 44229,
+  "lr-decay-iters": 44229,
+  "train-dataset-name": 'slim_pajama_100B_3_replay5',
+}
diff --git a/configs/datasets/train/slim_pajama_150B.yml b/configs/datasets/train/slim_pajama_150B.yml
@@ -0,0 +1,23 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/train_150B/ArXiv/ArXiv',
+    'data/slim_pajama/train_150B/Book/Book',
+    'data/slim_pajama/train_150B/C4/C4',
+    'data/slim_pajama/train_150B/Wikipedia/Wikipedia',
+    'data/slim_pajama/train_150B/Github/Github',
+    'data/slim_pajama/train_150B/StackExchange/StackExchange',
+    'data/slim_pajama/train_150B/CommonCrawl/CommonCrawl',],
+  "train-data-weights": [
+    4.576447650075095,
+    4.198505982426652,
+    26.62982374026485,
+    3.9945183507095225,
+    5.218824282422116,
+    3.372167199706489,
+    52.00971279439528
+  ],
+  "train-dataset-name": 'slim_pajama_150B',
+  "train-iters": 66342,
+  "lr-decay-iters": 66342,
+}
diff --git a/configs/datasets/train/slim_pajama_200B_1.yml b/configs/datasets/train/slim_pajama_200B_1.yml
@@ -0,0 +1,24 @@
+{
+  # or for weighted datasets:
+  "train-data-paths": [
+    'data/slim_pajama/tokenized_train_0-200B/ArXiv/ArXiv',
+    'data/slim_pajama/tokenized_train_0-200B/Book/Book',
+    'data/slim_pajama/tokenized_train_0-200B/C4/C4',
+    'data/slim_pajama/tokenized_train_0-200B/Wikipedia/Wikipedia',
+    'data/slim_pajama/tokenized_train_0-200B/Github/Github',
+    'data/slim_pajama/tokenized_train_0-200B/StackExchange/StackExchange',
+    'data/slim_pajama/tokenized_train_0-200B/CommonCrawl/CommonCrawl',
+  ],
+  "train-data-weights": [
+    3.4703977435152775,
+    3.904381603212791,
+    25.641950653802013,
+    3.804228253591696,
+    4.9994643949282045,
+    3.1815838172641993,
+    49.99799353368582,
+  ],
+  "train-iters": 88457,
+  "lr-decay-iters": 88457,
+  "train-dataset-name": 'slim_pajama_200B_1',
+}