6VecLM

6VecLM tokenises each nybble position as a word, feeds the sequence through a transformer encoder, and learns to autoregressively predict the next nybble. Sampling the trained model yields candidates that respect long-range dependencies captured in the attention layers.

• Reference: Placeholder citation (coming soon).

Train

rmap train --seeds seeds/hitlist.txt --output models/6veclm.bin six-veclm \
  --d-model 128 --n-head 8 --d-ff 512 --n-layers 6

Generate

rmap generate --model models/6veclm.bin --count 100000 \
  --output 6veclm.txt

Configuration

• --seed <u64> – RNG seed for parameter initialisation and sampling (default 42).
• --d-model <usize> – transformer hidden size (default 128).
• --n-head <usize> – number of attention heads (default 8).
• --d-ff <usize> – feed-forward network width (default 512).
• --n-layers <usize> – encoder stack depth (default 6).

Model notes

• Training leverages Candle; GPU acceleration is recommended for production-scale datasets.
• The generated model bundle packages tokenizer metadata and network parameters for reproducible inference.