Synthetic Persona Pretraining: Alignment from Token Zero
Paper • 2608.13482 • Published
revision string | seed int64 | seq_len int64 | gbs int64 | A_annotated int64 | K_canary int64 | N_total_seqs int64 | F_frontload int64 | train_steps int64 | compact_seqs int64 | train_tokens int64 | annotated_seq_share float64 | frontload_seq_share float64 | weights dict | model dict | lr dict | annotated_stats dict | built_at timestamp[s] |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
58460b5d9bb28975e5d38465cd1af7f7ec836a75 | 28 | 2,048 | 960 | 4,389,584 | 1,999 | 43,896,000 | 87,792 | 45,725 | 39,504,417 | 89,899,008,000 | 0.1 | 0.002 | {
"compact": 0.899954825041006,
"annotated": 0.09999963550209587,
"canary": 0.00004553945689812284
} | {
"hidden": 2048,
"layers": 25,
"heads": 32,
"kv_groups": 8,
"ffn": 7168,
"params": 1464178688,
"tokens_per_param": 61.39893220464632,
"target_params_3x_chinchilla": 1498316800
} | {
"peak": 0.0002,
"warmup_iters": 2000,
"wsd_decay_iters": 4572,
"style": "WSD linear->0"
} | {
"A": 4389584,
"K": 1999,
"doc_trunc": 600334,
"degenerate_n": 0,
"ann_content_tokens": 3134211759,
"refl_tokens_1p": 490435640,
"mean_insert_frac": 0.4958630476979269,
"labeled_row_count_prefilter": 4392128
} | 2026-08-18T01:17:35 |
op-spp-streams-v1 — tokenized Megatron streams (SPP-format pretraining corpus)
Tokenized Dolma v1.7 (ODC-BY)
subsample in Megatron IndexedDataset format (uint16, SmolLM2 tokenizer +
<assistant> extension from epfl-dlab/spp-training):
compact dense-packed 2049-token windows; annotated/canary one document
per window, EOD-padded. Built for a Synthetic-Persona-Pretraining-recipe run
(arXiv:2608.13482) — these files carry ONLY document tokens (raw public-corpus
text); the persona annotations ride in a separate private sidecar and are
inserted at training time. See corpus_meta.json for the exact run arithmetic.
- Downloads last month
- 34