Dolci-Think-SFT-translated

Machine translations of the Dolci-Think-SFT-32B dataset, produced with gemma-4-31B-it. The samples selected for translation are those where content_quality == "excellent" according to the propella annotations.

Columns

Each row is a translated conversation plus the result of a post-translation quality filter:

• id — source record id.
• messages — the translated conversation (list of {content, role}).
• filter_pass — true if the row passed every filter check (the recommended clean subset).
• reject_reasons — empty when filter_pass, otherwise any of: structure (broken <think>/turn mismatch), truncated (a trace/answer hit the generation token cap or has unbalanced delimiters), wrong_lang (target-language character fraction below 0.85), length (translation/source token ratio outside [0.60, 2.5]), degenerate (severe repetition).
• target_lang_frac — fraction of code-stripped prose characters in the target language.
• length_ratio — translated/source token ratio.
• src_tokens, trans_tokens — token counts (gemma-4-31B-it tokenizer).

The dominant rejection axis is truncated: some long reasoning traces exceeded the generation cap and were cut off mid-thought (these are complete-quality translations that simply ran out of token budget). Filter the dataset with filter_pass == true for the clean subset, or re-threshold using the raw columns.