pymlex/ru-bank-ie

Russian bank client information extraction benchmark with coverage-validated text-to-JSON pairs. Each example contains a chat-style client message, a gold BankClientExtraction JSON object, and a separate validation_json coverage justification. Fields may be null when absent from the source text.

Columns

• id — sample identifier
• reasoning — model planning before the client message
• text — client message used for evaluation
• gold_json — gold BankClientExtraction JSON
• validation_json — stage-3 coverage check with all_present, missing_fields, justification

Schema

Fixed Pydantic schema in schemas/bank_client.py with Russian field aliases.

Files

• test.jsonl — evaluation split, coverage-valid rows only
• test.csv — same data in CSV
• stage1_gold.jsonl — gold JSON profiles
• stage2_pairs.jsonl — text prompts from gold JSON
• stage3_validated.jsonl — all stage-2 pairs with validation_json
• benchmark/runs/ — per-model metrics CSV artefacts
• benchmark/results.json — aggregated benchmark scores
• benchmark/metrics.json — summary metrics JSON from GitHub results/
• benchmark/analysis.json — per-field and per-group breakdown
• benchmark/summary.csv — metrics table
• benchmark_assets/ — benchmark plots

Benchmark results

Evaluation on the full coverage-valid split ($N=368$) with three SLMs up to 2B parameters on NVIDIA RTX 5090. Metrics use normalised string comparison for phones, INN, passport codes, dates and case-folded text fields. The split is built from 500 stage-3 rows: only pairs with all_present=true and non-empty text are retained.

Model	SEM	Field F1	Entity F1	Null-field acc	Hallucination	Schema valid
Qwen3-1.7B	0.0	10.1	17.4	59.1	6.3	70.1
NuExtract-2.0-2B	0.8	39.2	56.1	81.2	5.9	100.0
LFM2-1.2B-Extract	0.0	5.7	11.9	56.4	7.4	100.0

• Evaluation split $N=368$ coverage-valid pairs from 500 stage-3 rows. Hardware: NVIDIA RTX 5090, bf16 inference.
• numind/NuExtract-2.0-2B leads on entity F1 (56.1%) and macro field F1 (39.2%). Three documents reach strict exact match. Schema validity is 100%.
• Qwen/Qwen3-1.7B reaches entity F1 17.4% with schema validity 70.1%. Identity fields extract better than passport blocks. Null-field accuracy is 59.1%.
• LiquidAI/LFM2-1.2B-Extract stays below 5.7% macro field F1. Passport and address groups are near zero.
• NuExtract field-group F1: IDs & contact 71.7%, work 63.8%, passport 54.5%.
• Address sub-fields with index, country and region stay at 0% F1 for NuExtract when gold omits them from client text.
• Metrics CSV and plots in this repository under benchmark/ and benchmark_assets/. GitHub mirror: ie-slm-bench results.

Plots

Generation

Dataset built with Qwen/Qwen3.5-4B and Outlines constrained decoding on Colab. Stage 1 generates 500 gold profiles with random field sparsity in $[0.2, 0.8]$. Stage 2 produces client messages split into reasoning and text. Stage 3 validates field coverage on text only.

bash scripts/generate_dataset.sh --n 500
bash scripts/finalize_and_push_dataset.sh --data-dir data/ru-bank-ie

Citation

@misc{zyukov2026ru_bank_ie,
  title={ru-bank-ie: Russian Bank Client Information Extraction Benchmark},
  author={Zyukov, Alexey},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/pymlex/ru-bank-ie}}
}

The project is under GPL-3.0 license.