nirantk/document-harness-evaluation

Talk 2 Harness Evaluation

Data behind the talk "Better Harness, Better Answers". One small model answered the same questions under different harnesses, and every answer, score and tool call is here. Variant: public.

Config	Rows	Content
questions	50	Question, gold answer, scoring rule, evidence pages. Splits follow the evaluation stages.
pages	71	Page image, converted Markdown, source URL, licence, the questions that selected the page.
answers	3625	Model, quantization, provider, seed, configuration, repeat, answer, score, tool calls, seconds, tokens, full tool trace.

Evaluation stages

• rag_misses, multi_hop: validation questions. The agent was changed after reading their traces.
• holdout: ten questions held back from the first agent, then used to change the harness. Treat them as validation data for any later revision.
• test: twenty questions selected before their first run. No harness change came from their results or traces. Fifteen ran first as holdout2; five more were added to match the twenty validation questions. Validation and test search the same 531 pages. In answers, corpus pages_validation holds the validation runs and pages_unseen the test runs; pages_multihop holds earlier validation runs on 222 of those pages, where the harness was tuned.

Retrieval and models

• Embedding model: qwen3-embedding:0.6b in Ollama (Qwen/Qwen3-Embedding-0.6B, 1024 dimensions), 512-token chunks with 50 overlap from LlamaIndex SentenceSplitter.
• Vector search: FAISS IndexFlatIP, an exact inner-product search over all chunks, top 5.
• BM25: bm25s. Hybrid fuses the top 20 of each ranking by reciprocal rank fusion (k = 60) and keeps 5.
• Page conversion: maternion/LightOnOCR-3:0.8b for pages with charts, images or little text, pymupdf4llm otherwise.
• Answer model: Qwen3.5-9B at temperature 0. Final runs use the 4-bit (fp4) weights on Darkbloom through OpenRouter, qwen/qwen3.5-9b, three runs. Earlier runs use the nvfp4 weights in Ollama, qwen3.5:9b, one run. Each row of answers names its model, quantization, hosted_provider and the seed sent; at temperature 0 the seed has little effect.
• Iterations: Trace Fixes is configuration merged_grep_revision2, Grep Token Reduction is merged_grep_revision3, and Grep Token Reduction + Thinking is merged_grep_revision3_thinking.
• BM25 agent (merged_bm25_loop), BM25 agent + Thinking (merged_bm25_loop_thinking) and Hybrid agent (merged_fusion_loop) are the grep agent with a ranked search tool, added after the test scores of the other configurations were known. Hybrid means embedding + BM25 with reciprocal rank fusion, top 5. text_bm25_loop is BM25 in the loop on PDF text.
• BM25 Token Reduction (merged_bm25_excerpt_fixed_count) puts BM25 search into the Grep Token Reduction agent, shows an excerpt of 600 characters around the matching words and has no turn limit. merged_bm25_excerpt_selected_count lets the model choose the passage count. merged_bm25_token_reduction and merged_bm25_selected_count are the first versions of these two. All four were tuned on the validation questions and added after the test scores of the others were known.

Files

files/ holds the PDF subset and Markdown of each document that permits redistribution. files/document_index.tsv lists every document with its source URL, including those not hosted here. pages.open_page_url opens a page directly.

Scoring

correct is re-scored with the current rules at export. match holds the rule for each question.

Variant notes

Every question, gold answer, scoring rule, evidence quote, model answer and score is here. Five XL-DocBench questions, four of them in test, come from an earlier release of that benchmark. Pages with redistribution other than permitted are excluded, because their source licences do not allow copies; source_url and open_page_url lead to them. Tool outputs and prompts that reproduce those documents read [omitted from the public variant: source licence].

Sources and attribution

• CUAD, The Atticus Project, CC BY 4.0, DOI 10.5281/zenodo.4595826. Contracts come from EDGAR.
• ViDoRe v3, annotations CC BY 4.0, arXiv 2601.08620. Each document has its own licence in the licence column. INSEE documents: "Source : Insee".
• XL-DocBench, Microsoft, arXiv 2608.00036. Questions and answers come from the benchmark; it lists links to documents and grants no redistribution rights for them.

This card is a practical summary and not legal advice.