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
import laya
agent = laya.load("goktugoguz/laya-multilingual-arc")
story = ("USDC moved from a wallet to a contract, amount 1 to 100 USDC (a small amount). "
"In the same transaction: tokens were swapped on an exchange. Protocol: Uniswap.")
agent.predict(story, {"q": {"type": "noul", "instructions": "About this Arc USDC transfer: Is this a swap?"}})
An MLX fp16 twin for Apple silicon (layad) is at goktugoguz/laya-multilingual-arc-mlx.
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
shapesentence, with the questionWhat kind of Arc transaction is this?and these options, verbatim and in this order. Order and wording are part of what was trained.
swap: tokens were swapped on an exchange or traded on a marketplacebridge: funds were sent to or arrived from another chain through a bridge, even if tokens were also swappedliquidity: pool liquidity changed, even if tokens were also swappedvault: funds were deposited into or withdrawn from a vault, or USDC was wrapped or unwrappedlending: a loan was opened, repaid or liquidatedsigned_payment: the payer signed an authorization and someone else submitted itpayment: a plain direct transfer, with nothing else happeningspam: zero or less than one cent of USDC moved and nothing else recognisable happened
- Viewer questions are asked of the
storysentence asnoulquestions, each prefixed withAbout 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.
| Questions | n | Balanced acc. (base) | Balanced acc. (fine-tune) | AUC (base) | AUC (fine-tune) |
|---|---|---|---|---|---|
| All countable | 59 | 0.786 | 0.966 | 0.877 | 0.982 |
| Written for this benchmark, before measuring | 36 | 0.748 | 0.950 | 0.850 | 0.971 |
| The radar's long-standing questions | 23 | 0.846 | 0.991 | 0.921 | 0.998 |
| Concepts the fine-tune was taught | 40 | 0.761 | 0.981 | 0.853 | 0.993 |
| Concepts it was never taught | 19 | 0.840 | 0.933 | 0.929 | 0.959 |
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
| Question | Taught | Base | Fine-tune |
|---|---|---|---|
| Did someone pay a fee in this transaction? (new) | no | 0.98 | 0.88 |
| Was a fee taken? | no | 1.00 | 1.00 |
| Was a token swap part of this transaction? (new) | yes | 0.99 | 1.00 |
| Is this junk with no real value? (new) | yes | 0.99 | 1.00 |
| Was a fee taken on a swap? (new) | no | 0.99 | 1.00 |
| Were rewards handed out or claimed? (new) | no | 0.99 | 1.00 |
| Were rewards claimed or paid out? | no | 0.99 | 1.00 |
| Is this a gasless signed payment? | yes | 0.99 | 1.00 |
| Was this swap done on Uniswap? (new) | no | 0.98 | 0.99 |
| Is this a KyberSwap trade? | no | 0.98 | 1.00 |
| Is USDC leaving Arc through a bridge? | yes | 0.98 | 1.00 |
| Did this go through CCTP? | yes | 0.99 | 1.00 |
| Was this routed through LI.FI? | no | 0.92 | 0.95 |
| Is this a Uniswap trade? | no | 0.91 | 0.94 |
| Was the trade on Aerodrome? (new) | no | 0.97 | 1.00 |
| Did this transfer happen inside a swap? (new) | yes | 0.95 | 0.97 |
| Was USDC minted? | yes | 0.97 | 1.00 |
| Did a contract send these funds? (new) | yes | 0.97 | 1.00 |
| Is this a swap? | yes | 0.96 | 1.00 |
| Was this a simple payment between two wallets with nothing else going on? (new) | yes | 0.89 | 0.96 |
| Is this a batch payout to many wallets? | no | 0.93 | 0.99 |
| Was the sender a smart contract account like ERC-4337? (new) | no | 0.56 | 0.63 |
| Was Circle's cross-chain protocol used? (new) | yes | 0.91 | 0.99 |
| Is this spam or dust? | yes | 0.92 | 1.00 |
| Was more than 100 USDC bridged out of Arc? (new) | yes | 0.78 | 0.86 |
| Was this bridged using LI.FI? (new) | no | 0.91 | 1.00 |
| Was this paid by signed authorization rather than by the sender's own transaction? (new) | yes | 0.87 | 1.00 |
| Is this a direct payment between two wallets? | yes | 0.86 | 0.99 |
| Did one transaction pay out to a lot of recipients? (new) | no | 0.85 | 1.00 |
| Did the USDC go into a contract? | yes | 0.83 | 1.00 |
| Is this smaller than a cent? (new) | yes | 0.82 | 0.99 |
| Was it sent by a smart account? | no | 0.82 | 1.00 |
| Is this less than one dollar? | yes | 0.79 | 0.99 |
| Was a contract on the receiving end? (new) | yes | 0.77 | 1.00 |
| Was this routed by OKX's DEX router? (new) | no | 0.59 | 0.83 |
| Did the Relay bridge carry this? (new) | yes | 0.75 | 1.00 |
| Is liquidity being added or removed? | no | 0.69 | 0.94 |
| Did KyberSwap route this trade? (new) | no | 0.73 | 0.99 |
| Did a contract pay a wallet as part of a swap? (new) | yes | 0.60 | 0.87 |
| Is this more than 100 USDC? | yes | 0.73 | 1.00 |
| Was less than a dollar moved in a swap? (new) | yes | 0.62 | 0.90 |
| Did a wallet receive the USDC? | yes | 0.71 | 0.99 |
| Did a regular wallet get these funds? (new) | yes | 0.50 | 0.79 |
| Did zero USDC change hands? (new) | yes | 0.67 | 1.00 |
| Is this transfer over 10,000 USDC? | yes | 0.67 | 1.00 |
| Did a wallet send USDC into a swap? (new) | yes | 0.65 | 1.00 |
| Was this part of decentralized finance? (new) | yes | 0.64 | 0.99 |
| Is this a DeFi transaction? | yes | 0.64 | 1.00 |
| Did wrapping or unwrapping of USDC happen here? (new) | yes | 0.63 | 1.00 |
| Did this move through a cross-chain bridge? (new) | yes | 0.64 | 1.00 |
| Was USDC created in this transaction? (new) | yes | 0.63 | 1.00 |
| Is the amount ten thousand USDC or more? (new) | yes | 0.62 | 1.00 |
| Was this USDC brought over from a different network? (new) | yes | 0.59 | 0.97 |
| Is this a transfer of a hundred USDC or more? (new) | yes | 0.61 | 1.00 |
| Did money cross from Arc to another network? (new) | yes | 0.59 | 1.00 |
| Did these funds arrive from another chain? | yes | 0.58 | 0.99 |
| Did a liquidity pool change size? (new) | no | 0.18 | 0.61 |
| Is this a bridge transfer? | yes | 0.58 | 1.00 |
| Is the value below one USDC? (new) | yes | 0.55 | 1.00 |
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. See NOTICE.
- Downloads last month
- 17
Model tree for goktugoguz/laya-multilingual-arc
Base model
convaiinnovations/laya-multilingual