BafoGPT-3B
This is gemma2-2b-base model continued-pretraining on the ChallengerSpaceShuttle/zulu-pretraining-dataset dataset.
This is the first iteration, on building IsiZulu models that can attain performance comparable to models that typically require millions of dollars to train from scratch.
🔍 Applications
This is the base model and has a context length of 8k. It can generate coherent Zulu text, one can finetune it based on instruction datasets.
⚡ Quantized models
🏆 Evaluation
🧩 Configuration
The code used to train the model can be found here: BafoGPT with the following training configuration.
model_name: google/gemma-2-2b
out_dir: pretrained_model/models
precision: bf16-mixed
initial_checkpoint_dir: google/gemma-2-2b
resume: false
data:
class_path: litgpt.data.LitData
init_args:
data_path: data
seed: 42
num_workers: 8
train:
save_interval: 1000
log_interval: 1
global_batch_size: 4
micro_batch_size: 1
lr_warmup_steps: 2000
max_tokens: 156800708
max_seq_length: 2048
tie_embeddings: false
max_norm: 1.0
min_lr: 4.0e-05
eval:
interval: 1000
max_iters: 100
initial_validation: false
final_validation: true
optimizer: AdamW
devices: auto
num_nodes: 1
tokenizer_dir: google/gemma-2-2b
logger_name: tensorboard
seed: 42
Architecture Config
{
"architectures": [
"Gemma2ForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"attn_logit_softcapping": 50.0,
"bos_token_id": 2,
"cache_implementation": "hybrid",
"eos_token_id": 1,
"final_logit_softcapping": 30.0,
"head_dim": 256,
"hidden_act": "gelu_pytorch_tanh",
"hidden_activation": "gelu_pytorch_tanh",
"hidden_size": 2304,
"initializer_range": 0.02,
"intermediate_size": 9216,
"max_position_embeddings": 8192,
"model_type": "gemma2",
"num_attention_heads": 8,
"num_hidden_layers": 26,
"num_key_value_heads": 4,
"pad_token_id": 0,
"query_pre_attn_scalar": 256,
"rms_norm_eps": 1e-06,
"rope_theta": 10000.0,
"sliding_window": 4096,
"torch_dtype": "float32",
"transformers_version": "4.42.4",
"use_cache": true,
"vocab_size": 288256
}