taozi555/llama3-Mirage-Walker-8b-v0.2

See axolotl config

axolotl version: 0.4.0

base_model: meta-llama/Meta-Llama-3-8B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: taozi555/bagel
    type: sharegpt
#  - path: jondurbin/cinematika-v0.1
#    type: text
  - path: MinervaAI/Aesir-Preview
    type: sharegpt
  - path: Norquinal/claude_multiround_chat_30k
    type: sharegpt
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./out
chat_template: alpaca

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

wandb_project: waifu-8b
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 2
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
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 100
eval_steps: 100
eval_table_size:
saves_per_epoch: 
save_steps: 100
save_total_limit: 20
debug:
deepspeed: /workspace/deepspeed.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

out

This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.7773

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: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
1.0419	0.0	1	1.1113
0.9179	0.07	100	0.8886
1.0123	0.14	200	0.8822
0.9106	0.21	300	0.8701
0.8992	0.28	400	0.8637
0.7915	0.35	500	0.8527
0.9123	0.42	600	0.8448
0.7849	0.49	700	0.8381
0.8381	0.56	800	0.8344
0.7652	0.63	900	0.8230
0.9006	0.7	1000	0.8167
0.8589	0.77	1100	0.8088
0.7635	0.84	1200	0.8016
0.7696	0.91	1300	0.7951
0.8476	0.98	1400	0.7879
0.6031	1.03	1500	0.8063
0.5386	1.09	1600	0.8065
0.5298	1.16	1700	0.8015
0.5736	1.23	1800	0.7979
0.5761	1.3	1900	0.7939
0.5576	1.37	2000	0.7917
0.4814	1.44	2100	0.7879
0.5146	1.51	2200	0.7842
0.4577	1.58	2300	0.7832
0.4821	1.65	2400	0.7806
0.6088	1.72	2500	0.7782
0.5113	1.79	2600	0.7785
0.5861	1.86	2700	0.7779
0.4885	1.93	2800	0.7773

Framework versions

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