disjointness_scope string | eval_windows int64 | format string | format_version int64 | index_sha256 string | kind string | max_anchor_row_fraction float64 | max_token_id int64 | max_train_anchor_draws int64 | max_training_rows int64 | mean_response_len float64 | mean_reward float64 | mean_weight float64 | n_episodes int64 | n_supervised_tokens int64 | n_tokens int64 | n_weighted_tokens int64 | numpy_version string | permutation_epoch int64 | permutation_seed int64 | permutation_tail_start int64 | source_episode_indices list | source_episode_indices_sha256 string | source_manifest_sha256 string | source_n_episodes int64 | source_window int64 | split_method string | supervised_token_count_basis string | tokens_sha256 string | weights_sha256 string | source_anchor string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
active_training_schedule_indices | 512 | raw | 1 | 36f030a78f7d1ea441e1f163d4905149fd24ac1a349b753437eb1bd4d3d16d69 | midtrain_schedule_holdout | 0.178926 | 50,256 | 732,881 | 4,096,000 | 321.216797 | 0 | 1 | 512 | 164,463 | 164,975 | 164,463 | 2.2.6 | 0 | 1,235 | 732,881 | [
110568,
134449,
536378,
160275,
162339,
91183,
170485,
420924,
525656,
609738,
709788,
3418,
670502,
38442,
299830,
57491,
521563,
587091,
136760,
77601,
449829,
520839,
239082,
254713,
274927,
139494,
85210,
127748,
536163,
485143,
474243,
494667,
588... | 2566079c114ebf6354b784ab4ba8e1ff75f8b5bae7e70adc0388440ad6bb53e8 | dce5b30e3e26ef8b1a30e58894eec63ebb59f21eefb9910fb6f7ddb857a7f42e | 733,393 | 1,025 | unused_epoch0_permutation_tail | count(token_weight > 0) | 97d67aded85aa7de098f963e1b945bebf0e025ef7a29d0e59df1c6050550aa49 | aa2b75bedd99807f597548405bae925039f26b08abcdc9954975ae407ba775cc | ../midtrain-240m |
Unified Training: midtraining data
Tokenized midtraining data of the real-world (0.76B) experiments in Training Foundation Models with a Single Algorithm. The models are in the model repo sfanm/unified-training.
| Folder | Paper item | Description |
|---|---|---|
realworld/midtrain-240m |
Sec. 5.2, App. D.1-D.3, D.5, E | 240M-token midtraining corpus with 24 components (web text, mathematics, code, question answering, reasoning traces) |
realworld/midtrain-code-removed |
App. D.4 | corpus variant without its three designated code sources (180M tokens) |
realworld/midtrain-code-replaced |
App. D.4 | corpus variant whose code allocation is execution-tested Python answers in chat format with assistant-only loss (240M supervised tokens) |
realworld/midtrain-val-holdout |
App. E | 512 held-out windows for the trainer's midtraining validation loss |
realworld/heldout-loss-slices |
App. D.1 | held-out documents of the 24 components for the held-out corpus loss |
Format
The first four folders use the trainer's round format: tokens.bin (int32 token ids), weights.bin (float32 per-token loss weights), index.npy (offset and length of each episode), manifest.json (counts, token totals, component weights, hashes) and an empty DONE marker. heldout-loss-slices holds one <component>.npz per component (tokens, offsets) and a manifest.json with the component weights. All token ids use the GPT-2 BPE tokenizer of the models.
Download one folder: huggingface-cli download sfanm/unified-training --repo-type dataset --include "realworld/midtrain-240m/*" --local-dir $DATA_DIR. The manifest of midtrain-val-holdout refers to ../midtrain-240m, so download both into the same directory.
License and citation
Released under the Apache-2.0 license. Derived from allenai/dolma3_dolmino_mix-100B-1025 (ODC-By 1.0); midtrain-code-replaced also contains Python answers from nvidia/OpenCodeInstruct (CC BY 4.0). The training corpora were decontaminated against the evaluation benchmarks. Citation: to be added. This card was prepared with AI assistance (Claude, an Anthropic model).
- Downloads last month
- 107