Matched scratch transaction/history position comparison
Both fresh MLMs and all eight downstream checkpoints are included; no variant is selected from test results. Common weights use paired fresh random initialization with seed 17 and eager attention. Every stage uses AdamW at learning rate 1e-4 and MPS, with neural CPU fallback disabled.
MLM training uses 20 epochs, physical batches of 512 windows, accumulation of 8, effective batches of 4,096 windows and drop_last=true. Each MLM makes 200 updates and 1,600 backwards over 819,200 used window exposures, with 1,060 omitted exposures from reshuffled incomplete groups. History training uses ten block epochs. Classification uses fifteen epochs of full 16-account batches, omitting 11 training accounts per new shuffle. Validation and test retain every account. The complete campaign contains 8,460 optimizer updates and 11,260 backwards.
Read experiment-report.md, audit.json, tracking-verification.json and review-ledger.md for all measured results, clipping, matched corruption, exact source fingerprints, conditional uncertainty and limits. The bundle includes six final optimizer/RNG states; trusted local PyTorch resume files are distinct from safe model exports. The source directory and execution-source.json bind the implementation. Generated per-model cards correct the generic training-output description without altering original local files. upload_manifest.json lists immutable uploaded-file checksums.
Requires the custom implementation from https://github.com/willynilly-fr/evm.
Load pretraining/{absolute,rotary}/model using evm_bert.model.load_mlm_model.
Load each {profile}-{architecture}-s17/model or finetune/model using
evm_bert.history_model.HierarchicalTransactionModel.from_pretrained and the raw-byte
transaction/history APIs. Stock AutoModel/text pipelines do not implement rotary
or history inputs. Pin the verified Hub revision recorded by the publication report.
The dense rotary encoder adapts the released positional rule, including value and output rotation. It retains dense BERT and separate K/V projections, not DeepSeek's full compressed/sparse architecture. Training uses at most 256 tokens per window; an explicit 2,048-position cap establishes no robust long-context quality guarantee. Kaggle FLAG labels historical accounts, not individual fraudulent transactions. Single-seed results on 54 test accounts do not establish deployment performance.