jeff-adapter-aml

Anti-money-laundering review. Applies an institution's monitoring policy to a customer's last 30 days - clear, investigate, escalate or block.

A LoRA adapter for jeff-base v1.3, a small open decision model (a fine-tune of Qwen3.5-0.8B). You send a situation (the state) and questions with named options; Jeff returns a calibrated probability for every option from one forward pass, with no generated text to parse. One Jeff server loads the base once and any number of adapters beside it; each request picks an adapter by name ("model": "aml").

Adapter page, with the full data card: jeffhub.ai/adapters/aml.

Results

On this adapter's held-out test sets, never trained on, scored three ways on the same rows: the untrained model Jeff is built from, the Jeff v1.3 base alone, and the base with this adapter. Questions have 2 to 8 options. As of 2026-10-05. All adapters

Each cell: accuracy · calibration error (ECE; lower is better, 0 is perfect).

By group (test set test)

By group (test set test-amlworld-v1)

With llama.cpp (GGUF)

The same test, through llama.cpp: the base GGUF (mstrasser/jeff-base-gguf) plus this adapter's LoRA GGUF (mstrasser/jeff-adapter-aml-gguf), with the temperature refitted for each format. Running Jeff with llama.cpp

Not measured yet for v1.3: calibration charts, the commonest confusions and accuracy per answer.

Source of these numbers: results/sources/v1.3/new-adapters.table.json in the JeffHub repository, also collected in jeffhub-v1.3.json.

When to use it

• You have a written transaction-monitoring policy (warning signs, thresholds, high-risk and blocked countries and parties) and want each customer review sorted first.
• You want the policy's own answer with a calibrated probability, so analysts start with the unsure cases.

When not to use it

• You want the model to find laundering your policy does not describe. It applies the rules you give it; on raw transactions without a written policy (IBM's AMLworld cross-test) it scores 60.0%, below the base alone.
• You need a final decision without a person. Reports to the authorities have legal consequences.
• Your activity summaries look very different from the training format. Test on your own data first.

How to use it

The adapter runs with Jeff's server, on the main branch of firelex/jeff, on the jeff-base v1.3 base.

git clone https://github.com/firelex/jeff && cd jeff
uv sync --no-default-groups --extra lora          # add --extra cuda on NVIDIA GPUs, --extra mac on Apple silicon
uv run --no-default-groups hf download mstrasser/jeff-base --revision v1.3 --local-dir checkpoints/jeff-base
uv run --no-default-groups hf download mstrasser/jeff-adapter-aml --revision v1.3 --local-dir adapters/aml
JEFF_CHECKPOINT=checkpoints/jeff-base JEFF_ADAPTERS=adapters/ PORT=8765 \
  uv run --no-default-groups jeff-serve          # on a Mac, add JEFF_BACKEND=mlx

Every folder in adapters/ is served under its folder name; add or replace adapters while the server runs with curl -X POST http://localhost:8765/v1/adapters/reload. Each adapter records the exact base it was trained on, and the server refuses an adapter trained on a different one, so this adapter loads only on jeff-base v1.3 (a v1.2 adapter does not load on v1.3). For llama.cpp, use mstrasser/jeff-adapter-aml-gguf.

Request format

State (the situation), in this order:

Questions:

• suspicious (choice): Whether the activity is suspicious under the policy. Options: yes, no
• action (choice): Which action the policy requires. Options: clear, investigate, escalate, block

Rules:

• Ask one question per request; each was trained with its own fixed instructions.
• Keep the institution and the customer profile the same across reviews, and put the activity last, so the unchanging part can be prepared in advance.
• The request format may still change before release.

General rules for every request: the request format guide.

Example

The request below is also in this repository as example.json.

{
  "model": "aml",
  "state": {
    "institution": "Meridian Crypto Exchange is a crypto exchange based in Italy. It serves these customer groups: private crypto investor, business accepting crypto payments. Clear means the review is closed with a short note; investigate means an analyst opens a case and reviews the customer's documents; escalate means the case goes to the money-laundering reporting officer, who decides on a report to the authorities; block means the payment is stopped and the account frozen until compliance decides.\n\nMonitoring rules (all amounts in euro):\nEach review covers one customer's last 30 days. Transfers between the customer's own accounts and wallets (listed in the customer profile) are ignored by every rule.\nWarning signs:\nW1 Structuring: 3 or more deposits made with cash vouchers bought at partner kiosks each from 2,400.00 up to just under the reporting threshold of 3,000.00.\nW2 Rapid pass-through: the incoming money totals at least 1,000.00 and at least 90% of it is sent out again (by any payment or cash withdrawal) within 72 hours of arriving.\nW3 Many senders: money comes from more different senders than the customer profile expects per month.\nW4 Many recipients: money goes to more different recipients than the customer profile expects per month.\nW8 High-risk country: a payment to or from a country on the high-risk list that the customer profile does not list as an expected country.\nHigh-risk countries: Algeria, Bulgaria, Cameroon, Lebanon, Mali, Nigeria, Philippines, South Sudan, Venezuela.\nBlocked countries: Belarus, North Korea, Syria.\nBlocked parties: Field Summit EOOD (Bulgaria); Ember Nova Company (Bolivia); Coral Alder Ltd (Malta).\nActions (exactly one applies):\nblock: any payment to or from a blocked country or a blocked party, whatever else the activity shows.\nescalate: two or more different warning signs, or W2 (Rapid pass-through) even on its own.\ninvestigate: exactly one warning sign, other than W2.\nclear: no warning sign and no blocked country or party.",
    "customer": "Customer: Kenji Okafor (individual), customer since November 2019.\nSegment: private crypto investor. Occupation or business: shop owner investing in crypto.\nExpected incoming: about 620.00 a month.\nExpected senders: up to 8 different senders a month. Expected recipients: up to 6 different recipients a month.\nExpected countries: Italy, Lebanon.\nOwn accounts and wallets: bank account ending 5014 at another bank; own hardware wallet 0xb84f...182c.",
    "activity": "Review period: 1 June 2026 to 30 June 2026 (30 days). Amounts in euro.\nPayments:\n01 Jun 09:12 received 46.76 by crypto deposit from wallet 0x576b...5f0d at FalconCoin (Italy)\n10 Jun 17:34 sent 376.32 by crypto withdrawal to wallet 0x3292...217b (private wallet, owner and country unknown)\n14 Jun 14:23 received 478.54 by crypto deposit from wallet 0x576b...5f0d at FalconCoin (Italy)\nSummary (transfers between own accounts left out):\n- incoming: 2 payments, total 525.30, from 1 different senders\n- outgoing: 1 payments, total 376.32, to 1 different recipients\n- share of incoming money sent out again within 72 hours: 0%\n- counterparty countries: Italy"
  },
  "questions": {
    "action": {
      "type": "choice",
      "instructions": "A compliance analyst at the institution described above reviews this customer's last 30 days of activity. Which action does the institution's monitoring policy require?",
      "criteria": {
        "block": "Block: at least one payment goes to or comes from a blocked country or a blocked party, whatever else.",
        "investigate": "Investigate: exactly one warning sign applies, and the policy does not say to escalate that one alone.",
        "escalate": "Escalate: two or more different warning signs apply, or a sign that the policy says to escalate alone.",
        "clear": "Clear: no warning sign applies at all, and no payment involves a blocked country or any blocked party."
      }
    }
  }
}

curl -s localhost:8765/v1/systemone -H 'content-type: application/json' -d @adapters/aml/example.json

The answer holds a probability for each option of each question. A recorded response from the v1.3 adapter is not published yet.

Files

• adapter_model.safetensors, adapter_config.json: the LoRA weights (PEFT format);
• readout.safetensors: the adapter's own readout over the answer codes;
• decision_config.json: answer codes, temperature, prompt layout and the checksum of the base it was trained on;
• example.json: the example request above.

adapter_config.json and decision_config.json name the base as mstrasser/jeff-base, revision v1.3; the server checks the base by the checksum of its weights.

Training

• 1.3.0 (2026-10-03): First release, trained on Jeff v1.3 with the live-last prompt layout (run 0.8b-aml-v2-20261004-0713, data version 2).

Data card

Self-reported. The numbers come from the adapter’s own maintainers and have not been re-run by anyone else. What the levels mean

• Test set: not attached yet
• QA report: not available here yet; it will be added once sanitised

How the test set was held out. Whole simulated institutions are held out: the test's 117 families (institutions, and for the pattern question traced laundering attempts) never appear in training, so its policies and customers are new to the adapter. 5,120 test rows: 2,386 per policy question (suspicious, action) and 348 pattern rows.

Training data. Training data not published.

Training data: 40,054 rows (version 2 of the data, built 2026-10-04); development 1,064, calibration 1,136 and test 5,120 rows.

Training mixed in a replay sample of the Jeff base model's own training data: 4,005 rows, about 10% on top of the adapter's 40,054 (inherited from the v1.2 recipe as a precaution; its effect has not been measured).

What it is not: it follows the institution's written policy, not a general laundering detector. On an out-of-distribution cross-test (4,500 windows from IBM's AMLworld simulation, asking only whether any transaction is part of laundering, with no written policy to apply) the adapter scores 60.0% (ECE 0.244), below the base alone at 64.3%. That test is shown for information; it is not what the adapter was trained to do.

Every label is computed by code - the scenario sets the action first, then the written policy's rules are re-applied to the transactions shown, and the build stops if the two disagree. GLM never sets a label.

A second question, which laundering pattern a traced set of transfers forms (fan-out, fan-in, cycle and others), uses rows built from IBM's synthetic AMLworld data.

Data and licence

Adapter licence: Apache-2.0.

Qwen3.5-0.8B notice: these weights were modified from Qwen3.5-0.8B by the Jeff project: jeff-base is a fine-tune of Qwen3.5-0.8B, and this adapter was trained on top of it. Qwen3.5-0.8B is Copyright 2026 Alibaba Cloud and licensed under the Apache License, Version 2.0; a copy of that licence is in LICENSE.

To confirm: whether the CDLA-Sharing-1.0 terms of the pattern rows affect the adapter's licence or only the data.

It was trained on:

• Synthetic institutions, customers and transactions. Licence: Released with the adapter under Apache-2.0 (made for this adapter) · Made by GLM 5.3 (own hardware), for some texts; everything else by code

  Simulated by code with fixed seeds; all names are invented. Some policy introductions, customer profiles and activity texts are written or reworded by GLM 5.3 and checked by code.

• IBM Transactions for Anti Money Laundering (AMLworld), laundering-pattern rows. Licence: Community Data License Agreement - Sharing - Version 1.0 (CDLA-Sharing-1.0); shared data, including modified data, must keep the same terms (open, but shared or changed data must keep the same terms) · Not made by a model

  Altman et al., "Realistic Synthetic Financial Transactions for Anti-Money Laundering Models", NeurIPS 2023 Datasets and Benchmarks. Fully synthetic; no real people or accounts.

Limitations

• Tied to jeff-base v1.3. It will not load on any other base or version; the server checks the base weights' checksum.
• Jeff chooses between the options you give it. It does not write text or reason in several steps.
• Calibration was fitted on this adapter's own calibration rows. On very different data, check it again.
• Everything listed under When not to use it above.

Links

• Adapter page: jeffhub.ai/adapters/aml
• Base model: mstrasser/jeff-base (revision v1.3)
• LoRA GGUF for llama.cpp: mstrasser/jeff-adapter-aml-gguf
• What changed in v1.3: release notes
• Code and server: github.com/firelex/jeff

Jeff is an independent project. It uses the same request format as Jev but is not affiliated with or endorsed by TypeSafe, the makers of Jev.