dataset string | split string | num_parquet_files int64 | tokenizer_repo_id string | tokenizer_kind string | tokenizer_local_path string | rendering string | conversation_field string | eos_token string | format string | num_documents int64 | total_tokens int64 | bin_bytes int64 | loss_masking string | prefix string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
allenai/Dolci-Think-SFT-32B | train | 156 | Qwen/Qwen3.5-0.8B | INSTRUCT (has chat_template.jinja) | /home/aiscuser/tmpnfs/models/Qwen3.5-0.8B-Instruct-tokenizer | tokenizer.apply_chat_template(messages, tokenize=False) then batch encode add_special_tokens=False; one Megatron document per conversation | messages (list<struct<content,role>>); assistant content carries inline <think>...</think> reasoning | <|im_end|> (id 248046); template appends trailing newline | Megatron indexed dataset (.bin/.idx), dtype int32 (vocab ~248k) | 2,253,684 | 24,180,699,916 | 96,722,799,664 | NOT embedded. Full rendered conversation tokens (prompt+assistant) are stored. Assistant-only loss masking must be applied at train time (no per-token mask in .bin/.idx; repo has no SFT masking tooling). | Dolci-Think-SFT-32B-q35instruct_text_document |
No dataset card yet
- Downloads last month
- 20