Model Details
Table with columns: Parameter, Value| Parameter | Value |
|---|
| Parameters | 32M |
| Architecture | LlamaForCausalLM |
| Hidden size | 384 |
| Layers | 16 |
| Attention heads | 6 (2 KV heads, GQA 3:1) |
| Head dim | 64 |
| MLP intermediate | 1024 |
| Vocab size | 16,384 |
| Context length | 2,048 |
| Tied embeddings | Yes |
Training
- Data: 40B tokens of FineWeb
- Tokenizer: Custom 16k vocab BPE trained on FineWeb
- Optimizer: AdamW (betas 0.9/0.95, weight decay 0.01, grad clip 1.0)
- Learning rate: 3.5e-3 with cosine decay, 1000 warmup steps
- Batch size: 524k tokens/step (64 micro × 4 grad accum × 2048 seq)
- Precision: bfloat16
- Hardware: 8× H100
Checkpoints
53 checkpoints are available as branches, log-spaced from step 1 to step 9536 (~40B tokens):
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("timaeus/dodecahedron-32m-v2")
model = AutoModelForCausalLM.from_pretrained("timaeus/dodecahedron-32m-v2", revision="step5108")
Available steps: 1, 2, 3, 4, 5, 6, 7, 8, 10, 12, 14, 16, 19, 22, 25, 30, 35, 40, 47, 55, 65, 75, 88, 103, 121, 141, 165, 193, 225, 263, 308, 359, 420, 491, 574, 671, 785, 917, 1072, 1253, 1465, 1713, 2002, 2340, 2735, 3198, 3738, 4369, 5108, 5970, 6979, 8158, 9536
Purpose
The Dodecahedron family is designed as small, well-characterized reference models for developmental interpretability and SLT research. Having densely-checkpointed models enables studying learning dynamics, phase transitions, and the geometry of the loss landscape throughout training.
