Laya Multilingual for Arc USDC transfers

A fine-tune of convaiinnovations/laya-multilingual (322M parameters), the base Laya decision model, at revision 1720e3e3. It is tuned to read the sentences Arc Radar writes about USDC transfers on Arc, Circle's stablecoin chain. Each transfer gets one of eight lanes and answers to yes/no questions that viewers ask about it.

On a benchmark of 66 English questions over 20,000 live transfers, the fine-tune raises balanced accuracy from 0.786 to 0.966. It does better than the base on 57 of 59 questions and worse on 1. An independent auditor checked 51 transfers on chain, drawn from those where the base and an earlier fine-tune (run 3) disagreed. It found our labels right in 51 of 51, this model right in 48 of 51 and the base right in 0 of 51.

Usage

Served on Apple silicon by layad:

LAYAD_MODEL=goktugoguz/laya-multilingual-arc-mlx layad serve

or loaded directly with laya-mlx:

import laya_mlx
agent = laya_mlx.load("goktugoguz/laya-multilingual-arc-mlx")

This is the fp16 MLX conversion of goktugoguz/laya-multilingual-arc (PyTorch). Over the validation rows, the two give the same lane every time. Their probabilities differ by at most 0.006.

What it reads

It reads one English sentence per transfer, never addresses. The sentences come from the radar's summarize() (source), and the model is only measured on them. Other text is out of its domain.

• Lane is asked of the shape sentence, with the question What kind of Arc transaction is this? and these options, verbatim and in this order. Order and wording are part of what was trained.

1. swap: tokens were swapped on an exchange or traded on a marketplace
2. bridge: funds were sent to or arrived from another chain through a bridge, even if tokens were also swapped
3. liquidity: pool liquidity changed, even if tokens were also swapped
4. vault: funds were deposited into or withdrawn from a vault, or USDC was wrapped or unwrapped
5. lending: a loan was opened, repaid or liquidated
6. signed_payment: the payer signed an authorization and someone else submitted it
7. payment: a plain direct transfer, with nothing else happening
8. spam: zero or less than one cent of USDC moved and nothing else recognisable happened

• Viewer questions are asked of the story sentence as noul questions, each prefixed with About this Arc USDC transfer: .

Example story: USDC moved from a wallet to a contract, amount 100 to 10,000 USDC (a medium amount). In the same transaction: funds were sent across chains through a bridge (out of Arc). Protocol: CCTP.

Benchmark

20,000 transfers, collected after this model was chosen. 66 questions: the radar's 27 English benchmark questions and 39 new ones, committed before either model answered them. A question counts when at least ten transfers answer it yes and ten no (59 do). Answers are read at each question's own yes line, the line the radar uses.

On the radar's own 1,200-transfer fixture, lanes agree with the audited fact table 99.8% (base) and 100.0% (fine-tune). The radar's question gate refuses 0 and 0 of 37 answerable questions, and turns away 8 and 14 of 14 nonsense ones.

Every question

The data, the per-question results and the audit are in the dataset repo goktugoguz/arc-usdc-laya-bench.

Training

• Data. 15,000 live Arc transfers plus 816 synthetic ones for facts the stream rarely carries. Both go through the radar's own sentence code. That gives 17,806 labelled questions over the sentences, from a question bank with held-out wordings, topics and languages, plus lane rows. A hand audit of 60 labels found all 60 backed by the sentence alone.
• Replay. 5,142 questions from outside the bank were trained toward the base model's own answers (learning without forgetting). Without them, fine-tuning pushed answers to untaught questions toward "no".
• Run. Full fine-tune with the token-embedding table frozen, soft cross-entropy on targets smoothed to 0.95/0.05, 3 epochs (laya-multilingual-arc-20261005-4). Trained on an M4 Pro Mac mini (MPS), under a supervisor that paused training whenever the live radars' model slowed.
• Blend. The released weights are 50% fine-tune and 50% base (WiSE-FT), which kept most of the gain and restored untaught questions.
• The lane temperature is the base's; the yes/no temperature was refit on validation.

Limits

• Measured on Arc USDC transfers only. Other chains, tokens or sentence formats are unmeasured.
• "Spam or dust" means any transfer under one cent, zero included, whatever else the transaction did.
• The yes line of a question is drawn from the radar's probe transfers. For a few untaught questions it lands in the wrong place even when the ranking is right.
• English only. Turkish was not a target; Spanish, German and Russian wordings were held out and measured.

License and attribution

Apache-2.0, as is the base model. Built on convaiinnovations/laya-multilingual by Convai Innovations. Converted with laya-mlx. See NOTICE.