baobabtech/decision-models-evaluation-docs

Decision models × evaluation documents

1,420 publicly published international development evaluation reports (about 40 organisations), with document labels and excerpt tags. Document labels come in two sets: silver labels from GLM-5.3-Flash (default columns) and the original pipeline labels (*_pipeline). Built for baobab-tech/decision-models-experiments: experiment 01 (zero-shot many-option classification) and 02 (fine-tuning).

Reports were converted from PDF with Docling and labelled by an LLM ingestion pipeline. The dataset is built by build_evaluation_docs_dataset.py from baobabtech/evalexplorer-data revision 5315eab; label names and definitions come from the pipeline's taxonomy and its document and excerpt prompts (eval-explorer commit 6ae90f7 for the excerpt prompts).

from datasets import load_dataset
docs = load_dataset("baobabtech/decision-models-evaluation-docs", "documents", split="test")
excerpts = load_dataset("baobabtech/decision-models-evaluation-docs", "excerpts", split="test").filter(lambda r: r["eval_sample"])
taxonomy = load_dataset("baobabtech/decision-models-evaluation-docs", "taxonomy", split="train")

Configs

Config	Rows (train / validation / test)	One row per
documents	1,148 / 138 / 134	report: first pages + document labels
excerpts	157,302 / 18,000 / 15,834	finding, recommendation or methodology excerpt + tags
taxonomy	334 (train)	label code

Splits are by document (80/10/10 on a hash of document_id), shared by both configs.

documents

Column	Contents
document_id, title	Report ID and title
text	First pages, exactly as the classify_codes config fed them to classifiers: 2 pages for reports under 10 pages, otherwise 5; cut at 24,000 characters. Median 1,935 tokens
n_chars, truncated	Length before truncation; whether it was cut (92 of 1,420)
evaluation_approach, evaluation_type, temporality	Silver labels from GLM-5.3-Flash (high reasoning effort, codes with one-line definitions, full first pages): one code or null. Test nulls: 24 / 10 / 37 of 134
themes, countries	Silver labels: themes (1–4 of 18 codes) and ISO 3166-1 alpha-2 countries
*_pipeline	The ingestion pipeline's labels (Gemini 2.5 Flash, gpt-oss-120b, Qwen 3 235B): evaluation_approach_pipeline, evaluation_type_pipeline, temporality_pipeline, themes_pipeline, countries_pipeline, regions_pipeline
label_source_pipeline	ai, or manual for the 36 documents whose pipeline labels were corrected by hand

excerpts

Column	Contents
excerpt_id, document_id, type, section_category, page	Identity and position
text	Verbatim excerpt; median 34 words, p90 107
themes (22 codes), regions (17), countries (198)	Pipeline labels, findings and recommendations. Tagged with the whole section as input
methods (24 codes)	Pipeline labels, methodology excerpts
eval_sample	true for the fixed experiment-01 test sample: 300 findings, 150 recommendations, 150 methodology, seed 0

taxonomy

field, code, label, definition (document fields only, from the document classification prompt), definition_excerpts (themes and methods, from the excerpt extract-and-classify prompts; longer theme definitions), region (countries only), in_documents, in_excerpts. The two flags mark the codes that occur in the silver or pipeline labels; experiments ask over those sets.

Labels

• Silver (default): GLM-5.3-Flash relabelled every document through HF Inference Providers. It returns null more often than the pipeline (test: approach 18%, temporality 28%).
• Pipeline (*_pipeline): the ingestion pipeline's LLM output; 36 document classifications corrected by hand.
• Pipeline labels score 76.2 mean field score against silver on test (exact match 9.0%).
• Excerpt tags are pipeline output only.
• Codes seen in fewer than 20 documents were dropped upstream.

Baselines on the same test split

Earlier Baobab Tech classifier runs on documents test (134), mean field score 0–100, against pipeline labels / against silver:

Model	Zero-shot	Fine-tuned (pipeline / silver)
Gemma 4 26B-A4B	70.1 / 72.9	84.4 / 80.3
Qwen3.5 4B	67.1 / 66.2	84.7 / 77.8
Gemma 4 E4B	72.3 / 72.1	83.0 / 76.8
GLiNER2.5 base	45.4 / 45.2	58.4 / 57.2

Licence

The reports are publicly published by their organisations, which keep their rights; check each report's terms before reusing its text beyond research. Labels, splits and taxonomy are released by Baobab Tech. No labels in this dataset are human gold labels except the 36 hand-corrected documents (label_source_pipeline == "manual"); all others are LLM output.