6GAN

6GAN tokenises IPv6 addresses, trains a generator/discriminator pair under reinforcement learning, and samples the learned policy to emit candidates that mimic the structural distribution of the seed set.

• Reference: Placeholder citation (coming soon).

Train

rmap train --seeds seeds/hitlist.txt --output models/6gan.bin six-gan \
  --num-classes 4 --embedding-dim 200 --hidden-dim 200

Generate

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

Configuration

• --seed <u64> – RNG seed used for both training initialisation and model metadata (default 42).
• --num-classes <usize> – number of generator heads trained for multi-pattern output (default 4).
• --embedding-dim <usize> – size of the token embedding layer (default 200).
• --hidden-dim <usize> – width of the LSTM layers (default 200).

Model notes

• Training leverages the Candle backend and currently targets CPU; large datasets may require patience compared to classical TGAs.
• Serialized models store aggregated metrics (pretrain, discriminator, reward) for reproducibility but do not yet export raw network weights.