I'll explain more about this model when I've found the optimal checkpoint for its use case
it's been full fine-tuned on Sandevistan.
Here is my Axolotl config (thanks to fizz and empti):
base_model: meta-llama/Meta-Llama-3-8B
load_in_8bit: false
load_in_4bit: false
strict: false
load_in_8bit: false
load_in_4bit: false
strict: false
datasets:
- path: Kquant03/Sandevistan_Reformat
type: customllama3_stan
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/out
max_steps: 80000
fix_untrained_tokens: true
sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
wandb_project: Pneuma
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
gradient_accumulation_steps: 16
micro_batch_size: 8
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00001
max_grad_norm: 1
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false
gradient_checkpointing: unsloth
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true
eval_sample_packing: false
plugins:
- axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true
hub_model_id: Replete-AI/L3-Pneuma-8B
hub_strategy: every_save
warmup_steps: 10
evals_per_epoch: 3
eval_table_size:
saves_per_epoch: 3
debug:
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
bos_token: "<|begin_of_text|>"
eos_token: "<|end_of_text|>"
pad_token: "<|end_of_text|>"
tokens:
This is the WandB loss for this section of the fine-tune:
