edededdy/fineweb-edu-grounded-qa

FineWeb-Edu Grounded Q&A

10,663 question–answer pairs, each grounded in a FineWeb-Edu passage that's included alongside it. They were generated to fine-tune ShallowSeek-mini, a tiny DeepSeek-style MoE language model, so that its SFT data would ask about text the model had actually seen during pretraining.

How it was made

1. Passages: 3,837 documents sampled at random from the FineWeb-Edu sample/10BT slice that ShallowSeek-mini was pretrained on (236,544 documents, 286M tokens). Only documents of at least 150 tokens were eligible, and each was truncated to its first 6,000 characters. That truncated text is exactly what's in passage.
2. Generation: a teacher model, Qwen3.8-9B (MLX 4-bit), thinking disabled, was asked for 3 question–answer pairs per passage, under these rules:
   • each question must make sense on its own, without seeing the passage;
   • each answer must be fully supported by the passage and stand alone, in 1–4 sentences of at most ~100 words;
   • questions should vary (facts, explanations, definitions, comparisons);
   • if the passage has nothing substantive (navigation text, ads, link lists), return no pairs.
3. Filtering: replies were parsed as JSON. Documents that produced no pairs (163 of them) appear only as empty marker rows in the raw generation log and are excluded here. No other filtering was applied.

Fields

Field	Description
question	The generated question
answer	The generated answer (median 37 words)
messages	The same pair in chat format: [{"role": "user", ...}, {"role": "assistant", ...}]
passage	The FineWeb-Edu source text the pair was generated from (truncated as above)
doc_index	Index of the source document in the generation run; pairs from the same passage share it
refers_to_passage	true for the 237 pairs (2.2%) whose question or answer refers to "the document", "the text" and so on, despite instructions. They're fine for passage-grounded use, but you may want to drop them for closed-book chat SFT

Uses

• Chat SFT (messages): short, factual, single-turn answers. Filter on refers_to_passage == false for closed-book use.
• Reading comprehension / RAG (passage + question → answer): every answer is checkable against its passage.
• Studying grounding: comparing a model's closed-book answers against the passage-supported reference.

Caveats

• Synthetic: written by a 9B-parameter model and not human-verified. Most pairs are accurate to their passage, but expect some errors, oversimplifications and awkward phrasing.
• Inherited from the web: FineWeb-Edu is filtered educational web text, but passages can contain dated, regional or incorrect information, which the answers inherit.
• Time-relative wording: some questions concern a passage's own moment ("the recent eruption…"), which can be ambiguous out of context.
• English only.

Licence and attribution

The passages come from FineWeb-Edu (Hugging Face), licensed under ODC-BY 1.0, and this dataset is released under the same licence. The question–answer pairs were generated by Qwen3.8-9B; check that model's licence terms if they matter for your use. Generation code: gen_sft_data.py.