Dolci-Think-SFT-7B-q35instruct

Megatron-format tokenization of allenai/Dolci-Think-SFT-7B using the Qwen3.5-0.8B-Instruct tokenizer and its chat template.

Contents

156 shards, each stored as two aligned Megatron indexed datasets (one document per conversation, one sequence per document):

• train-XXXXX-of-00156.bin / .idx — token ids (int32)
• train-XXXXX-of-00156_loss_mask.bin / .idx — per-token loss mask (uint8, 0/1)

624 files total (~113 GB).

Tokenization

Each row's messages conversation is rendered with the instruct chat template (apply_chat_template, add_generation_prompt=False) and tokenized. The loss mask is computed on the final rendered token stream via the assistant header delimiter <|im_start|>assistant\n … <|im_end|>:

• assistant response tokens (content + terminating <|im_end|>) = 1
• user / system turns, chat headers and formatting = 0

Stats

• Conversations: 2,268,178
• Total tokens: 24,205,560,617
• Assistant (train) tokens: 23,738,671,972

Loading

from megatron.core.datasets.indexed_dataset import IndexedDataset
tokens = IndexedDataset("train-00000-of-00156")
mask   = IndexedDataset("train-00000-of-00156_loss_mask")