Rewrite mixture cards: trim per-shard dumps, mark superseded/current, fix source paths

mixtures/v1-by-docs/ratio_00actual/manifest.json

@@ -1,7 +1,7 @@
 {
   "chars_per_shard": 50000000,
-  "midtrain_source": "zachnorton03/think-midtrain/stage5/data",
-  "old_corpus_source": "jbduran/think-dataset-clean-1930s",
+  "midtrain_source": "zachnorton03/bart-midtrain/corpus",
+  "old_corpus_source": "jbduran/bart-dataset-v3",
   "old_index_oversample_factor": 1.25,
   "random_seed": 42,
   "ratio_pct_midtrain_actual": 0.1,

mixtures/v1-by-docs/ratio_12actual/manifest.json

@@ -1,7 +1,7 @@
 {
   "chars_per_shard": 50000000,
-  "midtrain_source": "zachnorton03/think-midtrain/stage5/data",
-  "old_corpus_source": "jbduran/think-dataset-clean-1930s",
+  "midtrain_source": "zachnorton03/bart-midtrain/corpus",
+  "old_corpus_source": "jbduran/bart-dataset-v3",
   "old_index_oversample_factor": 1.25,
   "random_seed": 42,
   "ratio_pct_midtrain_actual": 11.53,

mixtures/v1-by-docs/ratio_25actual/manifest.json

@@ -1,7 +1,7 @@
 {
   "chars_per_shard": 50000000,
-  "midtrain_source": "zachnorton03/think-midtrain/stage5/data",
-  "old_corpus_source": "jbduran/think-dataset-clean-1930s",
+  "midtrain_source": "zachnorton03/bart-midtrain/corpus",
+  "old_corpus_source": "jbduran/bart-dataset-v3",
   "old_index_oversample_factor": 1.25,
   "random_seed": 42,
   "ratio_pct_midtrain_actual": 25.12,

mixtures/v2-by-tokens/ratio_21/README.md

@@ -1,130 +1,28 @@
-# Mixed corpus v2 -- ratio_21 (21% midtrain by tokens)
-
-Blend of [`zachnorton03/think-midtrain/stage5/data`](https://huggingface.co/datasets/zachnorton03/think-midtrain/tree/main/stage5/data)
-(midtrain) and [`jbduran/think-dataset-clean-1930s`](https://huggingface.co/datasets/jbduran/think-dataset-clean-1930s/tree/main) (original
-pretraining corpus), produced by `midtraining_mix_repair.ipynb`. Text-only shards (`text` column),
-document boundaries preserved -- nothing is concatenated or split to force the ratio.
-
-**This supersedes `mixed/ratio_21`**, which targeted its ratio by document count and
-therefore under-delivered midtrain content by token. That folder is left untouched.
-
-## Composition (measured with the run's exact tokenizer, BOS counted per document)
-
-| | tokens | share |
-|---|---|---|
-| midtrain | 864,644,049 | 20.9642% |
-| pretraining | 3,259,747,341 | 79.0358% |
-| **train total** | **4,124,391,390** | |
-| validation (`shard_00065.parquet`) | 22,603,190 | 20.30% midtrain |
-
-- Target: **21.00% midtrain**; achieved **20.9532%** over the
-  first 4,026,531,840 tokens the run actually consumes
-  (-0.0468 pp), **20.9642%** over the whole upload.
-- The ratio is enforced by a ratio-tracking interleave, so it holds over *every* prefix, not just
-  the total.
-- Supply: 4,124,391,390 train tokens vs 4,026,531,840 consumed
-  (1.0243x, 3% margin requested).
-
-## Provenance
-
-- Tokenizer: `jbduran/think.nano/experiments/clean1930s-d24-r12-ctx4096-sssl-fulltok-v1/tokenizer` (`rustbpe`, vocab 32,768),
-  fingerprint `89eafb1654dd04803378868d31d8efe2801873391a4f3df074d7f58544b110bf`
-- Midtrain source: `zachnorton03/think-midtrain/stage5/data` @ `9ace24b8e16e57e38a1ea0b1f6d7cbf323bf9dbc`
-- Pretraining source: `jbduran/think-dataset-clean-1930s` @ `45225d95bc15f942be3b4b344738cea1f66e3de8`, shards
-  331..396 (66 shards, starting
-  70% through the corpus). Disjoint from the base run's shards
-  0..329 and from the other v2 mixture.
-- Generator: `think.nano-midtraining` @ `None`
-- Shuffle seed: `42` (windowed shuffle, 3-shard window for the
-  pretraining side, 3 for midtrain)
-
-## Repetition and separation
-
-- Pretraining documents: 20,361 unique, **no repeats** (0 repeated ids,
-  0 duplicate text hashes).
-- Midtrain documents repeat by design: 15,073 emissions of
-  11,352 unique documents =
-  **1.4318 effective epochs** by tokens (1.3278 by
-  documents), starting 0% into the shuffled midtrain order and wrapping.
-- Validation is a held-out shard (185 documents); its documents appear nowhere in
-  training (0 id overlaps, 0 hash overlaps).
-- Cross-source text overlap: 0.
-- Dedup at index time removed 0 pretraining duplicates and
-  0 documents that duplicated midtrain content.
-
-## Layout
-
-`mixed/v2/ratio_21/data/shard_NNNNN.parquet` -- 65 train shards (indices 0..64) plus the
-validation shard at index 65. Per-shard sidecars with every document's id, source,
-occurrence, chars, exact tokens and text hash live in `mixed/v2/ratio_21/shard_meta/`.
-
-| shard | split | docs | tokens | midtrain | sha256 |
-|---|---|---|---|---|---|
-| `shard_00000.parquet` | train | 491 | 63,565,052 | 20.26% | `007c8e679178c095...` |
-| `shard_00001.parquet` | train | 525 | 63,415,663 | 20.20% | `66907706aafe64d8...` |
-| `shard_00002.parquet` | train | 533 | 63,269,875 | 20.48% | `13b2109dc84e60f1...` |
-| `shard_00003.parquet` | train | 509 | 63,777,503 | 20.36% | `619c637226b3c3fe...` |
-| `shard_00004.parquet` | train | 463 | 63,762,674 | 20.01% | `a8c563706ccf2e97...` |
-| `shard_00005.parquet` | train | 478 | 62,754,825 | 21.08% | `1a14f51e2d85b0c6...` |
-| `shard_00006.parquet` | train | 496 | 62,728,726 | 20.04% | `9c5dd5d24c20cf2b...` |
-| `shard_00007.parquet` | train | 481 | 64,039,500 | 20.04% | `82f75a2f4314e29a...` |
-| `shard_00008.parquet` | train | 518 | 62,940,635 | 20.05% | `88e0e5fb3162304a...` |
-| `shard_00009.parquet` | train | 525 | 63,347,542 | 20.33% | `b5672343cb2eb4b7...` |
-| `shard_00010.parquet` | train | 492 | 64,092,652 | 20.92% | `fa33097b20f3a9c4...` |
-| `shard_00011.parquet` | train | 497 | 63,601,280 | 20.59% | `5824fb15bf4d4136...` |
-| `shard_00012.parquet` | train | 483 | 63,769,217 | 20.76% | `6ed66c8d78d8c39c...` |
-| `shard_00013.parquet` | train | 498 | 64,869,204 | 22.30% | `36fa14eccbbbb79d...` |
-| `shard_00014.parquet` | train | 493 | 63,600,023 | 19.04% | `e3705e149df1346f...` |
-| `shard_00015.parquet` | train | 467 | 63,761,332 | 20.23% | `0eabab741d449ed7...` |
-| `shard_00016.parquet` | train | 462 | 63,637,512 | 21.08% | `d9a38a89aec31b24...` |
-| `shard_00017.parquet` | train | 528 | 63,676,079 | 21.05% | `dd2a8e2b7881cd66...` |
-| `shard_00018.parquet` | train | 601 | 64,012,876 | 21.48% | `315703aacb421576...` |
-| `shard_00019.parquet` | train | 584 | 63,465,770 | 21.32% | `902cdc0fc15de1c1...` |
-| `shard_00020.parquet` | train | 602 | 64,133,666 | 21.44% | `ad269fa8d98db947...` |
-| `shard_00021.parquet` | train | 643 | 64,192,041 | 21.10% | `abe42e0582f28645...` |
-| `shard_00022.parquet` | train | 655 | 65,280,803 | 21.20% | `68e74d90816bdcf3...` |
-| `shard_00023.parquet` | train | 651 | 63,753,841 | 21.60% | `6be4fcb2e7de20f8...` |
-| `shard_00024.parquet` | train | 627 | 64,258,817 | 21.68% | `6e414e960cc2281c...` |
-| `shard_00025.parquet` | train | 653 | 63,911,704 | 21.37% | `a7959a9469686243...` |
-| `shard_00026.parquet` | train | 613 | 64,809,622 | 21.13% | `817b0eb9e6bfeec2...` |
-| `shard_00027.parquet` | train | 571 | 64,267,802 | 21.51% | `48127e2174bf23b2...` |
-| `shard_00028.parquet` | train | 682 | 63,767,353 | 21.27% | `0286e1def7ec5640...` |
-| `shard_00029.parquet` | train | 677 | 63,145,056 | 21.76% | `5d5df57162ea8849...` |
-| `shard_00030.parquet` | train | 647 | 64,041,891 | 20.95% | `0dfdbd074358eea3...` |
-| `shard_00031.parquet` | train | 651 | 63,503,908 | 21.21% | `ce2c67d1ec73d805...` |
-| `shard_00032.parquet` | train | 635 | 63,876,194 | 21.16% | `9b4333b8f55ee5ec...` |
-| `shard_00033.parquet` | train | 705 | 63,752,519 | 21.55% | `4a421932b7cef0c7...` |
-| `shard_00034.parquet` | train | 672 | 63,902,558 | 21.45% | `6eb37383b25e5a3a...` |
-| `shard_00035.parquet` | train | 622 | 64,184,114 | 21.39% | `1f8aecd3c419a8c8...` |
-| `shard_00036.parquet` | train | 591 | 63,568,929 | 21.71% | `9d592288be2a1709...` |
-| `shard_00037.parquet` | train | 612 | 65,674,223 | 21.89% | `47f693d78085cb15...` |
-| `shard_00038.parquet` | train | 596 | 63,925,737 | 21.84% | `8cd5b37131976f2d...` |
-| `shard_00039.parquet` | train | 554 | 65,127,687 | 21.76% | `624afed822e9f768...` |
-| `shard_00040.parquet` | train | 565 | 64,014,979 | 22.27% | `1bf45fc5083967c3...` |
-| `shard_00041.parquet` | train | 538 | 65,582,594 | 21.97% | `64addfb00800207e...` |
-| `shard_00042.parquet` | train | 596 | 64,208,906 | 21.69% | `5ea68bc43a9ce173...` |
-| `shard_00043.parquet` | train | 597 | 64,084,152 | 22.13% | `776096865c8f25f9...` |
-| `shard_00044.parquet` | train | 557 | 63,734,570 | 21.37% | `048552cb6408ef29...` |
-| `shard_00045.parquet` | train | 498 | 62,978,877 | 20.76% | `76ff4c2c170a022c...` |
-| `shard_00046.parquet` | train | 512 | 62,740,028 | 20.23% | `c4a3ab3a88f68380...` |
-| `shard_00047.parquet` | train | 503 | 62,482,007 | 20.77% | `a32514bd022fdafb...` |
-| `shard_00048.parquet` | train | 486 | 62,875,041 | 20.52% | `315b8e9beef02b9d...` |
-| `shard_00049.parquet` | train | 485 | 63,202,259 | 19.99% | `7e0c855df02a81bc...` |
-| `shard_00050.parquet` | train | 504 | 63,074,335 | 20.51% | `b1c15893dbde262d...` |
-| `shard_00051.parquet` | train | 448 | 62,718,046 | 20.51% | `f1ea6b7af7fb07aa...` |
-| `shard_00052.parquet` | train | 500 | 62,879,644 | 20.30% | `d0130b3a9e46815e...` |
-| `shard_00053.parquet` | train | 511 | 63,425,910 | 20.17% | `9e79d2af7c94c18a...` |
-| `shard_00054.parquet` | train | 454 | 64,336,239 | 20.29% | `15c937fb266fc672...` |
-| `shard_00055.parquet` | train | 522 | 63,814,224 | 20.50% | `07b672f161219cdd...` |
-| `shard_00056.parquet` | train | 465 | 64,682,836 | 21.36% | `ac8b71f2055c638b...` |
-| `shard_00057.parquet` | train | 519 | 65,318,823 | 19.75% | `2aaae494bded0e3b...` |
-| `shard_00058.parquet` | train | 443 | 63,338,196 | 21.57% | `9ff8cdc785fd7846...` |
-| `shard_00059.parquet` | train | 492 | 64,340,720 | 19.76% | `fa24841556754900...` |
-| `shard_00060.parquet` | train | 458 | 63,657,355 | 21.04% | `7c112c3fccd4857e...` |
-| `shard_00061.parquet` | train | 455 | 63,748,942 | 20.65% | `1829bf6bf5131f87...` |
-| `shard_00062.parquet` | train | 562 | 64,479,100 | 20.95% | `c04844414249c60e...` |
-| `shard_00063.parquet` | train | 616 | 64,019,962 | 21.87% | `ac5145ca41ac4a50...` |
-| `shard_00064.parquet` | train | 365 | 39,487,240 | 20.97% | `d8c384f3a8409b4e...` |
-| `shard_00065.parquet` | validation | 185 | 22,603,190 | 20.30% | `28adedf3d79576ab...` |
-
-Updated: 2026-08-08T01:51:14Z
+# Mixture v2 (by tokens) — `ratio_21`
+
+> **CURRENT.** Ratio computed by **token count**. This supersedes
+> `mixtures/v1-by-docs/ratio_12actual`, which targeted its ratio by document count and therefore
+> under-delivered midtrain content.
+
+Blend of [`corpus/`](https://huggingface.co/datasets/zachnorton03/bart-midtrain/tree/main/corpus) (midtrain) and [`bart-dataset-v3`](https://huggingface.co/datasets/jbduran/bart-dataset-v3) (pretraining corpus),
+produced by `midtraining_mix_repair.ipynb`. Text-only shards (`text` column); document boundaries
+preserved — nothing is concatenated or split to force the ratio.
+
+| | |
+|---|---|
+| Target (by tokens) | 21.00% midtrain |
+| **Achieved (by tokens)** | **20.9642%** (whole upload) |
+| Achieved over consumed prefix | 20.9532% (-0.0468 pp vs target) |
+| Midtrain epochs | 1.4318 by tokens / 1.3278 by docs |
+| Train documents | 35,434 (11,352 unique midtrain / 20,361 pretraining) |
+| Train shards | 65 (+ validation shard 65) |
+| Tokens per shard (target) | 64,000,000 |
+
+The ratio is enforced by a ratio-tracking interleave, so it holds over *every* prefix of the
+stream, not just the total.
+
+**Deduplication.** Stable document ids (`source:shard#row`) plus a SHA-1 of lowercased,
+whitespace-collapsed text. Cross-source duplicates: 0.
+Train/validation overlap: 0.
+
+Per-shard records are in `manifest.json`; per-shard metadata sidecars are in `shard_meta/`.

mixtures/v2-by-tokens/ratio_21/manifest.json

@@ -14,7 +14,7 @@
     "midtrain_pct_consumed_prefix": 20.9532,
     "midtrain_pct_whole_upload": 20.9642
   },
-  "dataset": "zachnorton03/think-midtrain/mixed/v2/ratio_21",
+  "dataset": "zachnorton03/bart-midtrain/mixtures/v2-by-tokens/ratio_21",
   "dedup": {
     "index_cross_source_duplicates": 0,
     "index_empty_documents_dropped": 0,
@@ -45,9 +45,9 @@
     "columns": [
       "text"
     ],
-    "data_prefix": "mixed/v2/ratio_21/data",
+    "data_prefix": "mixtures/v2-by-tokens/ratio_21/data",
     "n_train_shards": 65,
-    "shard_meta_prefix": "mixed/v2/ratio_21/shard_meta",
+    "shard_meta_prefix": "mixtures/v2-by-tokens/ratio_21/shard_meta",
     "tokens_per_shard_target": 64000000,
     "validation_shard_index": 65
   },

mixtures/v2-by-tokens/ratio_45/README.md

@@ -1,98 +1,28 @@
-# Mixed corpus v2 -- ratio_45 (45% midtrain by tokens)
-
-Blend of [`zachnorton03/think-midtrain/stage5/data`](https://huggingface.co/datasets/zachnorton03/think-midtrain/tree/main/stage5/data)
-(midtrain) and [`jbduran/think-dataset-clean-1930s`](https://huggingface.co/datasets/jbduran/think-dataset-clean-1930s/tree/main) (original
-pretraining corpus), produced by `midtraining_mix_repair.ipynb`. Text-only shards (`text` column),
-document boundaries preserved -- nothing is concatenated or split to force the ratio.
-
-**This supersedes `mixed/ratio_45`**, which targeted its ratio by document count and
-therefore under-delivered midtrain content by token. That folder is left untouched.
-
-## Composition (measured with the run's exact tokenizer, BOS counted per document)
-
-| | tokens | share |
-|---|---|---|
-| midtrain | 933,536,546 | 45.0018% |
-| pretraining | 1,140,908,221 | 54.9982% |
-| **train total** | **2,074,444,767** | |
-| validation (`shard_00033.parquet`) | 22,788,633 | 43.47% midtrain |
-
-- Target: **45.00% midtrain**; achieved **45.0559%** over the
-  first 2,013,265,920 tokens the run actually consumes
-  (+0.0559 pp), **45.0018%** over the whole upload.
-- The ratio is enforced by a ratio-tracking interleave, so it holds over *every* prefix, not just
-  the total.
-- Supply: 2,074,444,767 train tokens vs 2,013,265,920 consumed
-  (1.0304x, 3% margin requested).
-
-## Provenance
-
-- Tokenizer: `jbduran/think.nano/experiments/clean1930s-d24-r12-ctx4096-sssl-fulltok-v1/tokenizer` (`rustbpe`, vocab 32,768),
-  fingerprint `89eafb1654dd04803378868d31d8efe2801873391a4f3df074d7f58544b110bf`
-- Midtrain source: `zachnorton03/think-midtrain/stage5/data` @ `09259d82d3f3237de346a3718a4e07c5195ab824`
-- Pretraining source: `jbduran/think-dataset-clean-1930s` @ `45225d95bc15f942be3b4b344738cea1f66e3de8`, shards
-  425..448 (24 shards, starting
-  90% through the corpus). Disjoint from the base run's shards
-  0..329 and from the other v2 mixture.
-- Generator: `think.nano-midtraining` @ `None`
-- Shuffle seed: `42` (windowed shuffle, 3-shard window for the
-  pretraining side, 3 for midtrain)
-
-## Repetition and separation
-
-- Pretraining documents: 7,370 unique, **no repeats** (0 repeated ids,
-  0 duplicate text hashes).
-- Midtrain documents repeat by design: 18,289 emissions of
-  11,273 unique documents =
-  **1.5596 effective epochs** by tokens (1.6224 by
-  documents), starting 44% into the shuffled midtrain order and wrapping.
-- Validation is a held-out shard (212 documents); its documents appear nowhere in
-  training (0 id overlaps, 0 hash overlaps).
-- Cross-source text overlap: 0.
-- Dedup at index time removed 0 pretraining duplicates and
-  0 documents that duplicated midtrain content.
-
-## Layout
-
-`mixed/v2/ratio_45/data/shard_NNNNN.parquet` -- 33 train shards (indices 0..32) plus the
-validation shard at index 33. Per-shard sidecars with every document's id, source,
-occurrence, chars, exact tokens and text hash live in `mixed/v2/ratio_45/shard_meta/`.
-
-| shard | split | docs | tokens | midtrain | sha256 |
-|---|---|---|---|---|---|
-| `shard_00000.parquet` | train | 916 | 65,415,824 | 45.19% | `fcdba371f9e1067f...` |
-| `shard_00001.parquet` | train | 876 | 64,172,439 | 44.45% | `8f16f497123e01fd...` |
-| `shard_00002.parquet` | train | 926 | 64,760,283 | 44.91% | `67855bfb87617c46...` |
-| `shard_00003.parquet` | train | 931 | 64,145,664 | 45.04% | `c371ecdd0f563399...` |
-| `shard_00004.parquet` | train | 1,048 | 64,448,692 | 45.28% | `26d0e2294e01ce79...` |
-| `shard_00005.parquet` | train | 824 | 64,993,694 | 44.94% | `ff3d0a783428bd65...` |
-| `shard_00006.parquet` | train | 820 | 65,108,738 | 46.13% | `09135ea15e9cfb6b...` |
-| `shard_00007.parquet` | train | 827 | 65,265,459 | 46.37% | `fc7dbaf027b40f12...` |
-| `shard_00008.parquet` | train | 783 | 65,401,294 | 46.10% | `61698fb0c7ea619e...` |
-| `shard_00009.parquet` | train | 681 | 64,278,692 | 45.32% | `57d776980e39bed8...` |
-| `shard_00010.parquet` | train | 642 | 62,562,998 | 43.71% | `cc9ca64f0cca14cf...` |
-| `shard_00011.parquet` | train | 592 | 62,538,627 | 43.73% | `c21d52e8775d6fef...` |
-| `shard_00012.parquet` | train | 571 | 62,725,614 | 43.59% | `b6c9996e3b65c240...` |
-| `shard_00013.parquet` | train | 643 | 62,234,413 | 43.54% | `4aa3bf668f2b6be4...` |
-| `shard_00014.parquet` | train | 636 | 63,024,871 | 44.42% | `6f5c07082a5df5c1...` |
-| `shard_00015.parquet` | train | 593 | 63,166,862 | 44.61% | `770ed8021f9a0639...` |
-| `shard_00016.parquet` | train | 539 | 63,250,420 | 44.23% | `ae6d165a7518a415...` |
-| `shard_00017.parquet` | train | 619 | 63,458,126 | 44.66% | `f19a42e1be8059f7...` |
-| `shard_00018.parquet` | train | 834 | 64,205,415 | 44.83% | `52da013ce8bcda3d...` |
-| `shard_00019.parquet` | train | 839 | 64,223,853 | 45.61% | `926cb59b7ef363ce...` |
-| `shard_00020.parquet` | train | 934 | 64,381,159 | 45.32% | `2ee71f5333942218...` |
-| `shard_00021.parquet` | train | 950 | 64,789,383 | 45.44% | `5c41cf4834ec82d7...` |
-| `shard_00022.parquet` | train | 875 | 64,065,456 | 44.92% | `306fef259c932da6...` |
-| `shard_00023.parquet` | train | 981 | 64,436,670 | 44.94% | `fbf9b88b9337d202...` |
-| `shard_00024.parquet` | train | 929 | 63,797,344 | 45.12% | `3e4955e37a53160e...` |
-| `shard_00025.parquet` | train | 1,021 | 64,430,138 | 45.03% | `3da5904833f830c6...` |
-| `shard_00026.parquet` | train | 832 | 64,590,269 | 45.62% | `33464c9856d57263...` |
-| `shard_00027.parquet` | train | 786 | 64,797,561 | 45.52% | `3eee2f8816c0948b...` |
-| `shard_00028.parquet` | train | 846 | 65,114,197 | 47.08% | `d1227d080a014019...` |
-| `shard_00029.parquet` | train | 816 | 65,774,411 | 46.04% | `a3de1b0c284a79cd...` |
-| `shard_00030.parquet` | train | 699 | 63,907,666 | 44.88% | `2019dbe06e39bd43...` |
-| `shard_00031.parquet` | train | 644 | 62,309,023 | 44.07% | `d7786408788df96f...` |
-| `shard_00032.parquet` | train | 206 | 22,669,512 | 42.34% | `9c096c5e2160c03d...` |
-| `shard_00033.parquet` | validation | 212 | 22,788,633 | 43.47% | `2bd7c1e1334bd192...` |
-
-Updated: 2026-08-08T03:15:13Z
+# Mixture v2 (by tokens) — `ratio_45`
+
+> **CURRENT.** Ratio computed by **token count**. This supersedes
+> `mixtures/v1-by-docs/ratio_25actual`, which targeted its ratio by document count and therefore
+> under-delivered midtrain content.
+
+Blend of [`corpus/`](https://huggingface.co/datasets/zachnorton03/bart-midtrain/tree/main/corpus) (midtrain) and [`bart-dataset-v3`](https://huggingface.co/datasets/jbduran/bart-dataset-v3) (pretraining corpus),
+produced by `midtraining_mix_repair.ipynb`. Text-only shards (`text` column); document boundaries
+preserved — nothing is concatenated or split to force the ratio.
+
+| | |
+|---|---|
+| Target (by tokens) | 45.00% midtrain |
+| **Achieved (by tokens)** | **45.0018%** (whole upload) |
+| Achieved over consumed prefix | 45.0559% (+0.0559 pp vs target) |
+| Midtrain epochs | 1.5596 by tokens / 1.6224 by docs |
+| Train documents | 25,659 (11,273 unique midtrain / 7,370 pretraining) |
+| Train shards | 33 (+ validation shard 33) |
+| Tokens per shard (target) | 64,000,000 |
+
+The ratio is enforced by a ratio-tracking interleave, so it holds over *every* prefix of the
+stream, not just the total.
+
+**Deduplication.** Stable document ids (`source:shard#row`) plus a SHA-1 of lowercased,
+whitespace-collapsed text. Cross-source duplicates: 0.
+Train/validation overlap: 0.
+
+Per-shard records are in `manifest.json`; per-shard metadata sidecars are in `shard_meta/`.

mixtures/v2-by-tokens/ratio_45/manifest.json

@@ -14,7 +14,7 @@
     "midtrain_pct_consumed_prefix": 45.0559,
     "midtrain_pct_whole_upload": 45.0018
   },
-  "dataset": "zachnorton03/think-midtrain/mixed/v2/ratio_45",
+  "dataset": "zachnorton03/bart-midtrain/mixtures/v2-by-tokens/ratio_45",
   "dedup": {
     "index_cross_source_duplicates": 0,
     "index_empty_documents_dropped": 0,
@@ -45,9 +45,9 @@
     "columns": [
       "text"
     ],
-    "data_prefix": "mixed/v2/ratio_45/data",
+    "data_prefix": "mixtures/v2-by-tokens/ratio_45/data",
     "n_train_shards": 33,
-    "shard_meta_prefix": "mixed/v2/ratio_45/shard_meta",
+    "shard_meta_prefix": "mixtures/v2-by-tokens/ratio_45/shard_meta",
     "tokens_per_shard_target": 64000000,
     "validation_shard_index": 33
   },