shisa-ai/shisa-v1-llama3-8b.lr-5e6

The was part of some LR ablations. It's not bad but you should probably prefer 8e-6

I ran the tests for 2 runs just to try to lower variance. These are all using temp 0.2, min_p 0.1, freq penalty 0.5

Model	AVG Score	ELYZA100	JA MT-Bench	Rakuda	Tengu-Bench	JA Char %
shisa-v1-llama3-8b.lr-2e4	3.97	4.60	4.54	3.33	3.42	92.42%
shisa-v1-llama3-8b.lr-5e5	5.73	6.28	6.45	5.37	4.81	90.93%
shisa-v1-llama3-8b (2e5 avg)	6.33	6.51	6.66	6.68	5.48	91.51%
shisa-v1-llama3-8b.8e6	6.59	6.67	6.95	7.05	5.68	91.30%
shisa-v1-llama3-8b.5e6	6.42	6.33	6.76	7.15	5.45	91.56%
shisa-v1-llama3-8b.2e6	6.31	6.26	6.88	6.73	5.38	92.00%

• The 2e-4 and 5e-5 are definitely overtrained and perform significantly worse.
• 2e-5 is on the edge since weightwacher shows the embed as slightly overtrained for 2e-5, but NEFTune version is not
• 8e-6 performs the best, and 5e-6 also performed slightly better than 2e-5

See axolotl config

axolotl version: 0.4.0

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

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: llama3
datasets:
  - path: augmxnt/ultra-orca-boros-en-ja-v1
    type: sharegpt
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/lr-5e6

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

use_wandb: true
wandb_project: shisa-v2
wandb_entity: augmxnt
wandb_name: shisa-v1-llama3-8b.lr-5e6

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: linear
learning_rate: 5e-6

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
evals_per_epoch: 2
eval_table_size:
saves_per_epoch: 0
debug:
deepspeed: axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.00
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

outputs/lr-5e6

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

• Loss: 0.5020

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: 5e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.3951	0.0064	1	0.8645
0.891	0.5020	79	0.5705
0.8575	1.0040	158	0.5243
0.7296	1.4853	237	0.5079
0.7068	1.9873	316	0.4976
0.6618	2.4694	395	0.5020

Framework versions

• Transformers 4.40.2
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1