A 300M-parameter language model trained from scratch on deqing/addition_dataset as part of the Convergent Evolution project, which investigates how Fourier features emerge in LLM number embeddings.

Model details


Architecture	LLaMA-style Transformer (12 layers, 1024 hidden, 16 heads, GQA)
Parameters	~300M
Optimizer	AdamW
Data perturbation	addition-only data (e.g., '123 + 456 = 579')
Training data	deqing/addition_dataset
Context length	1024
Tokenizer	Llama 3 (128K vocab)
Batch size	512 sequences

Usage

python

from transformers import AutoModelForCausalLM

# Load final checkpoint
model = AutoModelForCausalLM.from_pretrained("deqing/convergent-llama-300M-adamw-addition")

Training dynamics

Intermediate checkpoints are saved as branches: tokens-200M, tokens-400M, ..., tokens-5.0B.

python

# Load intermediate checkpoint (e.g., at 1B tokens)
model = AutoModelForCausalLM.from_pretrained("deqing/convergent-llama-300M-adamw-addition", revision="tokens-1B")

Citation

Paper forthcoming.