MaatAI/african-history-knowledge-merged-sft-cleaned

African History Knowledge Merged SFT — Cleaned

A reproducible, format-cleaned version of MaatAI/african-history-knowledge-merged-sft, pinned to source commit 0a40eb041d85d59b86219641de0fd87786ee0f77.

Cleaning performed

• —Quarantined 13 training records: 12 have no final answer after a closing thinking tag, and one has ambiguous repeated closing tags. Their original text and source row indices are preserved in quarantine.jsonl, which is not a training split.
• —Restored a missing opening <think> tag in 12 retained conversations (8 train, 3 validation, 1 test), preserving their existing reasoning and final-answer text exactly.
• —Added an empty <think>\n\n</think>\n\n block before direct answers, to align them with reasoning-model chat templates. An empty block is formatting, not a generated reasoning trace.
• —No text was summarized, truncated, translated, or synthetically expanded. User and system messages are unchanged.
• —Original split membership and retained-row order are preserved. Validation/test formatting is normalized identically, with no records removed from those splits.

Schema and usage

The dataset retains a single messages column containing lists of {role, content} objects. Apply the target model's native chat template; preserve existing reasoning, and compute loss only on assistant tokens, including their reasoning and final answers. Do not add a second set of thinking tags.

The source examples were audited with the Qwen3.5 and LFM2.5 tokenizers. A 24,576-token training limit fits the retained examples in the current comparison setup. Token lengths depend on the tokenizer and chat template: remeasure them when changing models, and reject unintended truncation.

Validation and provenance

validation.json records successful checks for schema round trips, idempotent normalization, preservation of original message text, no exact normalized duplicates, and no exact cross-split overlap. provenance.jsonl maps every published row to its original split/index and content hashes. manifest.json pins the source revision and normalization script used by the sequential Qwen3.5-2B / LFM2.5-2.6B training job.

These are structural checks. They do not establish factual accuracy, reasoning quality, absence of near-duplicates, or absence of source-document overlap. Empty thinking blocks must not be counted as reasoning supervision; explicit retained nonempty tagged reasoning is sparse. The 13 quarantined examples are available for manual repair without guessing missing answers or reasoning boundaries.

Source and license

The source dataset combines African history, literature, current events, and question-answering datasets. This derivative retains its declared Apache-2.0 license and English/French metadata; source attribution remains available in the original dataset card and pinned manifest.