Three standalone EOS-focused cipher models
Transformer, original Mamba-1 and standard ungated DeltaNet. Their original from-scratch one-pass states initialize a separately registered matched finetuning experiment. Optimizers and phase counters reset; sampler positions continue. Parent and additional exposure are recorded separately. Hybrid is deferred.
EOS weighting, validation-only EOS-logit calibration, and generated validation quality select candidates. Decoding has no ciphertext-relative or token-count cutoff. A runtime watchdog reports unfinished generation as failure; it never fabricates EOS or successful truncated predictions.
All official validation/test rows were generated. Character scores remove punctuation and literal NUL while retaining spaces; word/edit metrics preserve references and remove predicted NUL. Raw diagnostic predictions are retained in the source repository. Official splits contain content overlap, and historical test scores were already seen; results are exploratory. No test score selects settings.
The custom Mamba Triton scan and the original model equations remain. Checkpoint metadata, complete phase histories, parent lineage, selected-checkpoint exposure and measured memory are included. best.safetensors is the inference artifact; last.pt is a checksum-bound PyTorch training-state serialization from this pinned release.
Logs: https://wandb.ai/aryama-murthy-iiit-hyderabad/anlp-a3-part1-newdata