manifest_cap10_pretokenized.json

{
  "source_parquet": "data/sft-v1-cap10/rst_sft_train.parquet",
  "tokenizer": "data/Qwen3.5-27B-tokenizer",
  "rows_in": 10578,
  "rows_out": 10578,
  "dropped": {
    "contract": 0,
    "too_long": 0,
    "no_trained_tokens": 0,
    "error": 0
  },
  "total_tokens": 99939485,
  "trained_tokens": 32402050,
  "trained_fraction": 0.3242,
  "max_seq_len": 32768,
  "mask_source": "slime/utils/mask_utils.py::gen_multi_turn_loss_mask_qwen3_5 (ported)",
  "schema": {
    "input_ids": "list[int] \u2014 the exact tokens, whole-conversation render",
    "loss_mask": "list[int] \u2014 1 = train on this token, 0 = context only"
  },
  "note": "loss_mask is aligned 1:1 with input_ids: mask[i] refers to token i, with no offset applied. To get `labels`, set labels[i] = input_ids[i] where loss_mask[i]==1 else -100, and do NOT shift \u2014 every HuggingFace CausalLM (and Liger's fused CE) shifts internally, so shifting here misaligns the supervision by one token. Shift only if you hand-write the cross-entropy against unshifted logits, in which case shift logits and labels together as usual."
}