tFINE-base-300m

An encoder-decoder (T5 architecture) pretrained with nanoT5:

• tokenizer: sentencepiece BPE w/ byte fallback, 48k vocab (from vocab scaling laws)
• data: fineweb-edu-dedup split of HuggingFaceTB/smollm-corpus
• context length: 1024 ctx

details

Detailed info, including training logs, configs, and checkpoints can be found under checkpoints/ in this repo.

Expand hyperparameter overview

1. Model:

   • Dropout rate: 0.0
   • Activations: silu, gated-silu
   • torch compile: true

2. Data processing:

   • Input length: 1024
   • MLM probability: 0.15

3. Optimization:

   • Optimizer: AdamW with scaling
   • Base learning rate: 0.008
   • Batch size: 120
   • Total training steps: 80,000
   • Warmup steps: 10,000
   • Learning rate scheduler: Cosine
   • Weight decay: 0.0001
   • Gradient clipping: 1.0
   • Gradient accumulation steps: 24
   • Final cosine learning rate: 1e-5

4. Hardware:

   • Device: RTX 4080
   • Precision: bfloat16, tf32

plots

training loss

Expand grad and weights L2 norm plots

grad norm

weights norm

---