dtufail/nuremberg-trials-corpus

Nuremberg Trials RAG Corpus

A preprocessed retrieval corpus built from the International Military Tribunal (IMT) proceedings at Nuremberg (1945–1946), structured for hybrid dense-sparse RAG pipelines. 46,325 chunks covering daily trial transcripts, prosecution documents, the final judgment, and supporting briefs.

This corpus underpins Nuremberg Scholar, a RAG system using BGE-M3 hybrid retrieval, bge-reranker-v2-m3 cross-encoder reranking, and Llama-3.1-8B-Instruct generation.

Corpus Statistics

Field	Value
Total chunks	46,325
Session transcript chunks	~39,600
Document/brief chunks	~6,700
Date range	November 1945 – October 1946
Chunk size	512 tokens max, paragraph-boundary splits
Overlap	50 tokens
Embedding model	BAAI/bge-m3 (1024-d dense + sparse lexical)

File Structure

index/
  dense.faiss         # FAISS IndexFlatIP — 46,325 vectors, 1024-d float32
  metadata.jsonl      # Per-chunk metadata (one JSON object per line, row-aligned with FAISS)
  sparse.jsonl        # BGE-M3 sparse lexical weights (19,349 unique tokens, 7.1M non-zero entries)
  chunks.jsonl        # Full chunk text + metadata

Chunk Schema

Each record in chunks.jsonl:

{
  "chunk_id":    "sessions::01-02-46::0001",
  "collection":  "sessions",
  "slug":        "01-02-46",
  "source_url":  "https://avalon.law.yale.edu/imt/01-02-46.asp",
  "date_iso":    "1946-01-02",
  "speaker":     "COL. STOREY",
  "page_number": 255,
  "chunk_index": 1,
  "total_chunks": 79,
  "token_count": 548,
  "body":        "COL. STOREY: If the Tribunal please...",
  "text":        "[Date: 1946-01-02 | Source: 01-02-46 | Speaker: COL. STOREY | ...] COL. STOREY: ..."
}

text is the retrieval-formatted string passed to BGE-M3 at embed time. body is the raw chunk text.

Usage

from huggingface_hub import snapshot_download
from pathlib import Path
import faiss, json

path = snapshot_download(
    repo_id="dtufail/nuremberg-trials-corpus",
    repo_type="dataset",
    allow_patterns=["index/*"]
)
index_dir = Path(path) / "index"

# Load FAISS index
index = faiss.read_index(str(index_dir / "dense.faiss"))

# Load metadata
metadata = []
with open(index_dir / "metadata.jsonl") as f:
    for line in f:
        metadata.append(json.loads(line))

print(f"Loaded {index.ntotal} vectors, {len(metadata)} metadata records")

Retrieval Pipeline

Built for a five-stage hybrid retrieval pipeline:

1. Dense retrieval — BGE-M3 query → FAISS inner product search (top-100)
2. Sparse retrieval — BGE-M3 sparse weights → inverted index lookup (top-100)
3. RRF fusion — Reciprocal Rank Fusion merges dense + sparse lists (top-25)
4. Reranking — bge-reranker-v2-m3 cross-encoder scores 25 candidates
5. Generation — Top-5 chunks → Llama-3.1-8B-Instruct (4-bit NF4 quantization)

Sources

Trial transcripts sourced from:

• Yale Avalon Project — Primary source for IMT transcripts
• Harvard Nuremberg Trials Project — Used to fill gaps where Yale transcripts were incomplete or missing

Original transcripts are public domain. This derived corpus (chunking, embeddings, metadata extraction) is released under CC BY 4.0.

Citation

@dataset{tufail2025nuremberg,
  author    = {Tufail, Daniyal},
  title     = {Nuremberg Trials RAG Corpus},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/dtufail/nuremberg-trials-corpus}
}