A 300M-parameter language model trained from scratch on FineWeb-Edu 10BT (~9.4B tokens, 1 epoch) 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	Muon (for 2D weights) + AdamW (for embeddings/bias/norm)
Data perturbation	block-diagonal attention mask (numbers cannot attend to context)
Training data	FineWeb-Edu sample-10BT (~9.4B tokens)
Context length	1024
Tokenizer	Llama 3 (128K vocab)
Batch size	512 sequences

Training dynamics

Intermediate checkpoints are saved as branches: tokens-200M, tokens-400M, ..., tokens-9.6B.

python

from transformers import AutoModelForCausalLM

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

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

Citation

Paper forthcoming.

deqing/llama-isolate-old

convergent-llama-300M-muon-isolate

Model details

Training dynamics

Citation

deqing/
llama-isolate-old