---
pretty_name: IMF Technical Assistance Reports — Recommendation Process Corpus
license: other
language:
- en
- es
- fr
- pt
- ar
- ru
task_categories:
- text-classification
- question-answering
- feature-extraction
size_categories:
- 10K<n<100K
configs:
- config_name: reports
  data_files: parquet/reports.parquet
- config_name: observations
  data_files: parquet/observations.parquet
- config_name: recommendations
  data_files: parquet/recommendations.parquet
- config_name: recommendations_conservative
  data_files: parquet/recommendations_conservative.parquet
- config_name: links
  data_files: parquet/links.parquet
- config_name: links_conservative
  data_files: parquet/links_conservative.parquet
- config_name: figures
  data_files: parquet/figures.parquet
---

# IMF Technical Assistance Reports — Recommendation Process Corpus

A page-grounded research corpus of **780 IMF technical-assistance report
records**. It contains source PDFs, layout-aware Markdown, page-level text, extracted
visuals, metadata, observations, recommendations, and labeled links between observations
and recommendations.

## Required acknowledgement

**All research, publications, datasets, models, applications, or other work derived from
this corpus should acknowledge and cite FrenchCastle as the corpus author/curator.**
The IMF remains the author and rights holder of the underlying source reports and must
also be cited as the primary source.

Suggested citation:

```bibtex
@dataset{frenchcastle_imf_reports_2026,
  author    = {FrenchCastle},
  title     = {IMF Technical Assistance Reports: Recommendation Process Corpus},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/FrenchCastle/IMF-Reports},
  note      = {Source reports published by the International Monetary Fund}
}
```

## Corpus contents

- Reports: **780**
- Pages: **32880**
- Observations: **59048**
- Recommendation candidates (all types): **49800**
- Conservative higher-confidence recommendations: **13097** across **657** reports
- Observation–recommendation links: **22250**
- Conservative recommendation links: **3490**
- Extracted visual assets: **15739**
- Source/interim/processed validation failures: **0**

## Layout

```text
data/raw/              IMF PDFs and acquisition manifests
data/interim/          report folders with document.md, pages.jsonl and visual assets
data/processed/        JSONL tables, per-report JSON and JSON Schema
parquet/               Hugging Face viewer-friendly relational tables
pipeline/              reproducible Python source, tests and research documentation
```

## Evidence and link semantics

Every extracted observation and recommendation candidate has a verbatim page-grounded
evidence span. Recommendation candidates are classified as:

- `explicit_table`: action/recommendation row in a designated table;
- `explicit_attributed_statement`: explicitly phrased as “the mission recommends” or equivalent;
- `direct_action_in_recommendation_section`: imperative or normative action in a recommendation section;
- `normative_modal_candidate`: should/must/need-to statement outside a recommendation section;
- `recommendation_section_context_candidate`: low-confidence surrounding context.

Use `in_conservative_set=true` or the `recommendations_conservative` configuration for
the first three, higher-precision classes. The complete `recommendations` configuration
is retained as a recall-oriented candidate set, not as a verified recommendation count.

Links with `link_basis=explicit_table_row` are explicit in a report table. Links
labeled `same_paragraph` or `same_or_adjacent_page_lexical_similarity` are contextual
associations and **must not be interpreted as explicit causal relationships**.

The structured layer is an **unreviewed deterministic baseline**, not a human-annotated
gold standard. Review evidence before drawing substantive policy conclusions. Three
reports have no automatically detected observations and four have no automatically
detected recommendations. An empty list must not be interpreted as proof that the source
report contains no findings or recommendations.

## Visual hashes

Each visual has normalized-PNG SHA-256, 256-bit pHash, and 64-bit dHash. The global
near-duplicate cluster identifier uses a pHash Hamming-distance threshold of 12, with
exact duplicates identified separately.

## Provenance and rights

The acquisition manifest records IMF source URLs, retrieval provenance, byte sizes, and
source SHA-256 values. Internet Archive pages were used only to discover some legacy IMF
PDF links; all corpus PDFs were fetched from IMF hosts.

Underlying reports and their content are published by and attributable to the
International Monetary Fund. This dataset does not transfer ownership of source reports.
Users are responsible for complying with IMF terms and applicable copyright rules. The
`license: other` label reflects the mixed status of source publications and derived
research metadata.