KemiOm/poetry-meter-best

Task

• Input: one poetic line
• Output: stress pattern string (e.g., -+-+-+-+-+) Example:
• Input: Tired Nature's sweet restorer, balmy Sleep!
• Output: +-+-+-+-+-+ This is a meter-only labeling model, not a full poetry generator.

Training Data Format

• input: poetic line text
• target: stress/meter pattern string The model is trained to map each line directly to its metrical stress pattern.

Model Details

• Developer: KemiOm
• Base model: google/flan-t5-large
• Method: LoRA fine-tuning
• Library: Hugging Face Transformers + PEFT
• Language: English
• Domain: formal/historical poetry (including eighteenth-century material)

Best Checkpoint and Training Setup

• epochs: 5.0
• learning rate: 1e-4
• weight decay: 0.01
• warmup ratio: 0.1
• per-device train batch size: 8
• gradient accumulation steps: 1
• per-device eval batch size: 4
• max input length: 256
• max target length: 64
• LoRA rank (r): 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• LoRA target modules: q,v
• seed: 42
• precision flags: bf16=true, fp16 enabled

Intended Uses

• Automatic meter annotation for poetic corpora
• Meter diagnostics for generated verse
• Preprocessing/evaluation component in constrained poetry pipelines
• Educational tooling for stress-pattern analysis

Limitations

• Stress ambiguity in English can produce multiple plausible scansions.
• Archaic spellings and punctuation can affect tokenization and stress assignment.
• Performance may drop on modern slang, non-poetic text, or out-of-domain genres.
• Meter labels reflect dataset conventions and annotation assumptions.

Usage