See axolotl config

axolotl version: 0.4.0

base_model: winglian/llama-3-32k-merged
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
hub_model_id:  KolaGang/Red_Llama_32_base
hub_strategy: end
load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: Drewskidang/chatlaw
    type: sharegpt
    conversation: chatml
  - path: Drewskidang/tool
    type: sharegpt
    conversation: chatml
  - path: rxavier/economicus
    type: sharegpt
    conversation: chatml
  - path: KolaGang/mergers
    type: alpaca
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
eval_sample_packing: False
output_dir: ./out

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

wandb_project: swag_llama
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 2
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
flash_attn_cross_entropy: false
flash_attn_rms_norm: true
flash_attn_fuse_qkv: false
flash_attn_fuse_mlp: true

warmup_steps: 100
evals_per_epoch: 4
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero1.json # multi-gpu only
weight_decay: 0.1
fsdp:
fsdp_config:
tokens:
  - "<|im_start|>"
  - "<|im_end|>"

Red_Llama_32_base

This model is a fine-tuned version of winglian/llama-3-32k-merged on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6810

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 5
• gradient_accumulation_steps: 4
• total_train_batch_size: 40
• total_eval_batch_size: 10
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
0.8855	0.02	1	0.9452
0.7195	0.26	16	0.7678
0.6507	0.52	32	0.6943
0.6398	0.79	48	0.6700
0.5713	1.03	64	0.6622
0.5277	1.29	80	0.6616
0.5166	1.55	96	0.6582
0.5437	1.82	112	0.6500
0.3328	2.06	128	0.6977
0.2989	2.32	144	0.6900
0.2852	2.58	160	0.6821
0.2714	2.84	176	0.6810

Framework versions

• Transformers 4.40.0.dev0
• Pytorch 2.2.0+cu121
• Datasets 2.15.0
• Tokenizers 0.15.0