metadata
datasets:
- sade-adrien/redpajama_v2_sample_10M
MappingAdapter exact structure available in representation_mapping.py
Mapping "sentence-transformers/stsb-roberta-large"'s hidden representation to "mistralai/Mistral-7B-Instruct-v0.1"'s.
Training:
Steps: 114k
Gradient accumulation: 2
Batch size: 64
Warm-up steps: 100
Learning Rate: 3e-5 with linear scheduling
Eval steps: %8000
Training hours: ~98h
Eval hours: ~10h
Gradient updates: 57k
Train examples: 7.3M
Eval examples: 106k
Adapter: Decoder_dim (4096) → 4096 → LeakyRelu(.1) → Encoder_dim (1024)