bartdegoede/small-rag-answers

Short RAG answers for small language models

Questions about technical writing, each with the top 4 passages a small embedding model retrieved for it, and a short answer written from those passages only, or a refusal when the passages don't answer the question. It was made to fine-tune a 0.6B-parameter model for an in-browser "ask my blog" page on bart.degoe.de, where the model must answer in at most two sentences from retrieved passages, and say so when they don't cover the question.

Code, results and the full write-up of what each version did to the model: github.com/bartdegoede/small-rag-finetune.

Made with Claude

This dataset is synthetic. The questions were written by Claude Sonnet 5.5 and the answers by Claude Opus 5.5 (Anthropic), through the Message Batches API. Answers were filtered automatically (over two sentences, or mentioning "passages", was dropped) and read through by hand, but none were edited.

Configurations

config	examples (train / validation / test)	refusals	answers
v3 (default)	4,530 / 255 / 260	16%	one short sentence with the fact asked for, in the passage's own words
broad	1,199 / 77 / 76	11%	overview and "how do I" questions about a whole document, answered in one or two sentences
v1	4,529 / 256 / 259	13%	up to two dense sentences combining several facts

v3 and v1 share the questions and passages; only the answers differ.

What they did to Qwen3-0.6B (LoRA, one epoch, judged by Claude Opus 5.5 on the blog's 45 questions: 30 the blog answers, 15 it doesn't):

	faithful (of 30)	wrong answers	invented (of 15)
base model	17	9	2
v1	8	18	0
v3	16	3	0
v3 + broad	13	7	0

v3 is the one that worked: as often right as the base model, with a third of its wrong answers. A 0.6B model imitates the density of v1's answers without reading the passages as carefully, and the broad answers, which summarise several passages, made it garble overview questions. Both are kept for comparison.

Fields

Each row:

• messages: system, user (numbered passages and the question) and assistant turns, ready for chat fine-tuning (e.g. mlx_lm.lora, TRL's SFTTrainer).
• answer, refusal: the assistant turn, and whether it is the refusal "The blog does not seem to cover this."
• id, source, doc: where the question came from (doc is the document's slug in its source).
• kind: answerable (the passage the question was written from is among the 4), removed (that document was taken out of retrieval, so the expected answer is usually a refusal), near-miss (a related question the passages probably don't answer), and in broad, broad / broad-removed.
• question; gold_rank and gold_retrieved (v1, v3) or doc_retrieved (broad): whether retrieval found the source passage or document by itself. If it didn't, the source passage was put in at a random position.
• teacher: the model that wrote the answer.

The teacher decided whether to answer, so a near-miss or removed question that a neighbouring passage does answer has a real answer.

How it was made

1. Corpus. 2,193 documents from four openly licensed sources (below), split 90/5/5 into train/validation/test by document, so no document is in two splits.
2. Retrieval. Documents are chunked (600 characters, 120 overlap) and embedded with minishlab/potion-base-8M at 128 dimensions; each question gets the top 4 chunks. The passages are as noisy as they are in production.
3. Questions by Claude Sonnet 5.5 from one chunk (v1/v3), or from a document's title, headings and opening (broad).
4. Answers by Claude Opus 5.5 under the system prompt in messages, with extra style guidance per version.

Sources and attribution

The passages and questions are derived from these sources, each pinned to a commit. Text from the documents is quoted in the passages, and the answers reuse its wording.

source	repository (commit)	licence
go-blog	golang/website, _content/blog	CC-BY-4.0, © The Go Authors
rust-blog	rust-lang/blog.rust-lang.org, content	MIT OR Apache-2.0, the Rust project
rust-rfcs	rust-lang/rfcs, text	MIT OR Apache-2.0, the Rust project
peps	python/peps, peps	public domain / CC0-1.0

Changes: documents were converted to plain markdown, split into chunks and quoted as passages; nothing else was edited. The Rust passages remain available under MIT OR Apache-2.0; both licence texts are in licenses/. The dataset as a whole (the selection, questions and answers) is released under CC-BY-4.0.

Limitations

• English technical writing about Go, Rust and Python only. The system prompt says "blog" for every source.
• Synthetic and not corrected by hand: some answers will be wrong or refuse when they shouldn't.
• The refusal is one fixed sentence; a model trained on this will use exactly that wording.

Citation

@misc{degoede2026smallrag,
  author = {Bart de Goede},
  title = {Short RAG answers for small language models},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/bartdegoede/small-rag-answers}}
}