fwe_transformer_340m_lion_lr3e_4_b1_0_9_b2_0_99_20260506_010507

This is a 340M Transformer++ language model trained on FineWeb-Edu using the Lion optimizer, as part of the OpenTome optimizer benchmark.

Model Details

Field	Value
Architecture	Transformer++
Model Type	transformer
Scale	340M
Hidden Size	1024
Layers	24
Attention Heads	16
Vocab Size	32000
Max Sequence Length	8192
Training Dataset	FineWeb-Edu
Optimizer	Lion
Hyperparameters	lr=3e-4, β₁=0.9, β₂=0.99
Timestamp	20260506_010507

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "OpenRaiser/fwe_transformer_340m_lion_lr3e_4_b1_0_9_b2_0_99_20260506_010507",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "OpenRaiser/fwe_transformer_340m_lion_lr3e_4_b1_0_9_b2_0_99_20260506_010507",
    trust_remote_code=True,
)

Training

See train.sh for the full training command. Training was conducted on H200 GPUs using the FLAME framework.

Citation

If you use this model, please cite the OpenTome benchmark paper.