typhoon-ai/AMLworldCompactEval

AMLworld-Compact

Paper: Under review · Code & evaluation · Baseline results

AMLworld-Compact is an evaluation set of 6,021 cases for anti-money-laundering research. HT-Coreset selects these cases from the HI-Small and LI-Small file-order test partitions of IBM's synthetic AMLworld dataset. Each case contains a local graph extracted from the full source graph and serialised as text, a target transaction ID in metadata, an illicit/benign label, an optional laundering typology, and an importance weight for estimating full-split metrics.

The graph inputs used in the reported LLM runs are described in Evaluation limitations.

Configuration	Evaluation cases	Illicit cases retained	Full test edges	Reduction
HI-Small	3,753	1,251	1,015,669	271×
LI-Small	2,268	756	1,384,810	611×

The source 60/20/20 partition follows released CSV row order without timestamp sorting. Time ranges overlap across train, validation, and test. The released coreset draw uses sampling seed 0.

Both configurations contain a single test split. All illicit edges are retained; benign edges are sampled using supervised difficulty strata. The released set has one illicit case for every two benign cases: the smallest benign budget evaluated in the construction study, giving its largest tested reduction.

Each row carries case_id, center_edge_id, subset_index, label (with the equivalent illicit flag), typology and typology_id (0–7, or −1 when unannotated), ht_weight, n_transactions, n_tokens_approx, and typed_graph_text. Typology IDs 0–7 correspond to fan-out, fan-in, cycle, scatter-gather, gather-scatter, stack, bipartite, and random, in that order.

Quick start

pip install datasets scikit-learn

from datasets import load_dataset

ds = load_dataset("typhoon-ai/AMLworldCompactEval", "HI-Small", split="test")
case = ds[0]
print(case["case_id"])
print(case["typed_graph_text"])

Use "LI-Small" to load the other configuration.

The main LLM runs used typed_graph_text as the graph input; label, illicit, typology, and typology_id are for scoring. Task instructions and few-shot examples are in the code repository's prompt templates. ICL-FS and ICL-ZS share task instructions and graph inputs; ICL-FS adds eight illicit and four benign training demonstrations. Graphs are extracted over two hops, with at most 50 neighbours per expanded account per hop, counting incoming and outgoing neighbours together. For new target-level evaluations, use the target-marked inputs.

The dataset viewer may shorten long cells; loading the dataset returns the complete text.

The supplementary frontier-API probe uses 198 HI-Small cases with separate instructions and condensed graph inputs. Its ZS-Graph and ZS-Base variants both have no demonstrations; the few-shot variants also add account roles and computed structural features.

Target-marked inputs for new evaluations

extras/targeted_prompts_v1/ provides a separate input version for the same 6,021 targets. Each evaluation_prompt explicitly identifies the target, includes its transaction exactly once, and gives bank-qualified endpoints and elapsed time over the included graph. Outcome labels are excluded from the prompt. Target inclusion, transaction fields, endpoint presence, and time summaries are checked against the source records.

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="typhoon-ai/AMLworldCompactEval",
    repo_type="dataset",
    filename="extras/targeted_prompts_v1/HI-Small/test-00000-of-00001.parquet",
)
prompts = pd.read_parquet(path)
text_to_send = prompts.iloc[0]["evaluation_prompt"]

Use the complete evaluation_prompt directly. Join predictions to the original scoring table by configuration and case_id; case order, target IDs, and weights are retained. The original data/ tables and published LLM predictions still refer to the archived format. No model results are reported for this new input version. Edge counts are preserved by replacing a non-target edge when the target was absent; the revised text can have different token lengths. Context is still drawn from the full source graph, without a prospective time cutoff. The new instructions, target fields, endpoint names, time summary, and occasional edge replacements change the representation together. The manifest and schema document the checks and provenance.

Evaluate predictions

Report Horvitz–Thompson (HT)-weighted precision, recall, and F1 as the primary detection metrics. Given one binary predictions array in dataset row order (1 = illicit, 0 = benign):

from sklearn.metrics import precision_recall_fscore_support

precision, recall, f1, _ = precision_recall_fscore_support(
    ds["label"],
    predictions,
    average="binary",
    sample_weight=ds["ht_weight"],
    zero_division=0,
)
print(f"P={precision:.4%}  R={recall:.4%}  F1={f1:.4%}")

Join external predictions by (configuration, case_id) before scoring: case IDs repeat across configurations. Choose any decision threshold using separate validation data.

Both summaries use the same targets selected by HT-Coreset and the same predictions:

Summary	What it measures
HT-weighted (full test)	Estimates for the full source test partition, using ht_weight
Unweighted (subset)	Scores on the same released 1:2 subset with every weight set to one, used for diagnosis

For fixed predictions and positive inclusion probabilities, weighted confusion counts are unbiased estimates of full-test counts. Precision and F1 are ratios of those counts and can have sampling bias and variance. Recall is unchanged by weighting because every illicit edge is retained with weight one. Predicting every case illicit gives 50% unweighted F1, so that value alone does not establish useful detection.

The reported typology macro-F1 (TF1) covers 791 HI-Small and 174 LI-Small illicit cases with a known ground-truth typology, including missed detections. A benign verdict or missing predicted typology is scored as none. The macro average uses the union of true and predicted labels, including none or legitimate when present. An unannotated reference typology does not mean benign; those illicit cases remain in detection evaluation but are excluded from TF1. Published TF1 values were computed from archived, postprocessed labels; the companion code's current live parser does not reproduce every historical free-text fallback.

The evaluation guide provides commands to score the released ensemble, run LLMs, and evaluate saved predictions.

The code also provides a sampling-uncertainty analysis and a held-out-family study. The latter uses cached full-test GFP-booster predictions over 500 draws per configuration; it does not compare full-test and coreset LLM predictions.

Files

load_dataset() loads only the evaluation tables, about 31.7 MiB across both configurations.

Directory	Contents
data/<configuration>/	Test Parquet files with graph text, labels, and weights
extras/<configuration>/	Aligned scoring arrays, graph features, and case indices, in evaluation-table row order
extras/targeted_prompts_v1/	Separately versioned target-marked prompts, schema, and integrity manifest; no new model predictions
ml_baselines/	Supervised checkpoints under weights/ and full-test predictions under test_probs/, five seeds each

The code repository provides loaders and scoring examples for the arrays and checkpoints.

Models

GFP means Graph Feature Preprocessor. Each booster uses 73 GFP signals and 6 raw transaction attributes (79 inputs). GCPAL adds five line-graph-derived features, giving 84 inputs for that construction scorer. The primary ML reference averages LightGBM+GFP and XGBoost+GFP across five training seeds. Both are trained on the training portion of the file-order partition. Their probabilities are in extras/*/ensemble_probs_coreset.npy. The inherited thresholds are 0.80 for HI-Small and 0.48 for LI-Small; they were selected on the full test set for the original construction scorer and have not been retuned.

The frozen construction ensemble also included GCPAL+GFP, whose random fine-tuning split overlaps roughly 60% of the test edges. The original three-model probabilities remain in construction_probs_coreset.npy to reproduce the sampling design; GCPAL is excluded from the primary task comparisons. Each configuration's scoring_metadata.json distinguishes the two roles. The seven evaluated LLMs are listed in the code repository's baselines table.

Evaluation limitations

The reported LLM runs used the original graph text for all 6,021 cases. Target IDs are stored in center_edge_id metadata. The graph text does not explicitly mark the target, and the task template retains the literal <ID>. In 324/3,753 HI-Small and 297/2,268 LI-Small cases, the target transaction itself is absent after neighbour capping, including 128 and 181 illicit targets respectively. The Time span field counts distinct timestamps rather than elapsed time.

The original graph strings are retained to reproduce the reported runs. The LLM scores and trace analyses describe model behaviour under this input format and do not isolate input effects from reasoning errors. A separate target-marked version is available for new evaluations.

Uses

This synthetic benchmark supports research on transaction classification, graph-to-text prompting, and error analysis. Use the released test split for evaluation and separate data for training and tuning. Results describe the released AMLworld splits and do not establish performance on real banking transactions.

Licences and source

Derived from IBM AMLworld. The original transaction CSVs are not included.

• Data and features (data/, extras/): CDLA-Sharing-1.0.
• Supervised model parameters and outputs (ml_baselines/): MIT.

See NOTICE.md for attribution and licence scope.

Contact

Open a GitHub issue for questions about the dataset or the code.

Acknowledgments

This work was initiated at SCB 10X in late 2025, following the release of our earlier work, FinCoT. We are grateful to Oravee Smithiphol for introducing our team to Phume Ngampornsukswadi, whose interest in FinCoT led to this collaboration. We also thank the members of the Typhoon team for their support, and Duncan Halverson for contributing to the early stages of this work.

Citation

BibTeX

@misc{nitarach2026amlcompact,
  title  = {AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis},
  author = {Nitarach, Natapong and Ngampornsukswadi, Phume and
            Taveekitworachai, Pittawat and Nonesung, Surapon and
            Sirichotedumrong, Warit and
            Pipatanakul, Kunat},
  year   = {2026},
  note   = {Under review}
}

Please also cite AMLworld. Its citation is included in the companion README citation section.