MohammadKhosravi/llama3.1-8b-pure-pmt-ielts-4class

Pure PrefixMemory-Tuning (PMT) on IELTS 4-Class Dataset (B1-C2)

This experiment trains the uncompressed PMT architecture directly on IELTS Task 2 essays to evaluate CEFR generation boundaries (B1, B2, C1, C2) independent of the EFCAMDAT domain.

Architecture Details

1. 1.Uncompressed PMT: Full 4096*4096 $M$ matrices across 32 layers.
2. 2.Explicit CEFR Anchor: nn.Embedding(4, 4096) mapping directly to B1, B2, C1, and C2.
3. 3.Training Scale: Processed utilizing pre-defined train/test splits from the Hugging Face IELTS benchmark repository.

Training Performance

[
  {
    "epoch": 1,
    "train_loss": 3.0506,
    "val_loss": 2.7033,
    "val_ppl": 14.93
  },
  {
    "epoch": 2,
    "train_loss": 2.4405,
    "val_loss": 2.6086,
    "val_ppl": 13.58
  },
  {
    "epoch": 3,
    "train_loss": 2.042,
    "val_loss": 2.7409,
    "val_ppl": 15.5
  },
  {
    "epoch": 4,
    "train_loss": 1.3414,
    "val_loss": 3.0368,
    "val_ppl": 20.84
  }
]