Edit model card

shisa-v2 Base Model ablation

Per the Llama 3 Community License Agreement, the official name of this model is "Llama 3 shisa-v1-llama3-70b"

This is a fine-tune Llama 3 70B Instruct with the primary shisa-v1 dataset to improve Japanese language capabilities.

This model uses a LR of 8e-6 that slightly improves performance vs the initial 2e-5 tune (based on and validating predictive power of the the results of the Llama 3 8B LR ablations).

It also uses NEFTune, although the expected impact is neglible for this dataset.

While the 2e-5 model matched gpt-3.5-turbo performance, this 2e-6 version consistently edges it out, so I think it's fair to say that this model "beats" it.

While this is merely a test ablation on the road to shisa-v2, as of its release (mid-May 2024), it's the strongest commercially-usable open JA model benchmarked so far, so this model may be of general interest.

Performance

Measured using a fork of Lightblue's Shaberi benchmark framework:

Model Average ELYZA-tasks-100 MT-Bench Rakuda Tengu-Bench
gpt-4-turbo-2024-04-09 8.75 8.78 8.74 9.18 8.31
gpt-4o-2024-05-13 8.72 8.88 8.69 9.15 8.16
gemini-1.5-pro 8.58 8.58 8.93 9.20 7.61
claude-3-opus-20240229 8.55 8.64 8.58 8.75 8.23
CohereForAI/c4ai-command-r-plus 7.69 7.50 7.43 9.05 6.79
shisa-ai/shisa-v1-llama3-70b 7.30 7.34 7.67 8.15 6.04
gpt-3.5-turbo-0125 7.17 7.24 6.98 7.64 6.82
shisa-ai/shisa-v1-llama3-70b.2e5 7.17 7.16 7.45 7.98 6.09
karakuri-ai/karakuri-lm-8x7b-chat-v0.1 7.00 7.18 6.30 7.98 6.55
karakuri-ai/karakuri-lm-70b-chat-v0.1 6.84 6.86 6.43 7.85 6.23
lightblue/ao-karasu-72B 6.81 7.19 6.54 7.25 6.27
shisa-ai/shisa-v1-llama3-8b 6.59 6.67 6.95 7.05 5.68
microsoft/Phi-3-medium-128k-instruct 6.48 7.10 5.92 6.84 6.04
shisa-ai/shisa-v1-swallowmx-13a47b 6.17 6.48 6.07 7.11 5.03
lightblue/suzume-llama-3-8B-japanese 5.96 6.68 4.96 6.68 5.53
augmxnt/shisa-gamma-7b-v1 5.82 5.96 5.02 6.85 5.47
shisa-ai/shisa-v1-phi3-14b 5.77 6.28 5.26 6.55 5.01
shisa-ai/shisa-v1-gemma-8b 5.64 6.50 5.42 5.10 5.55
Rakuten/RakutenAI-7B-chat 5.58 5.92 4.60 6.58 5.24
lightblue/qarasu-14B-chat-plus-unleashed 5.20 5.58 4.74 5.46 5.01
shisa-ai/shisa-v1-mistral0.3-7b 5.11 5.64 6.10 3.83 4.86
cyberagent/calm2-7b-chat 4.76 4.90 3.58 5.75 4.81
mistralai/Mistral-7B-Instruct-v0.2 4.69 5.78 4.65 3.80 4.53
shisa-ai/shisa-v1-yi1.5-9b 4.63 5.98 4.28 3.26 5.00
augmxnt/shisa-7b-v1 4.50 4.63 3.95 4.89 4.53

Built with Axolotl

See axolotl config

axolotl version: 0.4.0

base_model: meta-llama/Meta-Llama-3-70B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

# doesn't work...
# hub_model_id: shisa-ai/shisa-llama3-70b-v1
# hub_strategy: end

use_wandb: true
wandb_project: shisa-v2
wandb_entity: augmxnt
wandb_name: shisa-llama3-70b-v1.8e6

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/basemodel-llama3-70b.8e6

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

neftune_noise_alpha: 5

gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: linear
learning_rate: 8e-6

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

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_ratio: 0.1
evals_per_epoch: 2
eval_table_size:
saves_per_epoch: 0
debug:
deepspeed: axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

outputs/basemodel-llama3-70b.8e6

This model is a fine-tuned version of meta-llama/Meta-Llama-3-70B-Instruct on the None dataset. It achieves the following results on the evaluation set:

  • Loss: 0.4440

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: 8e-6
  • train_batch_size: 2
  • eval_batch_size: 2
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 16
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 64
  • total_eval_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 87
  • num_epochs: 3

Training results

Training Loss Epoch Step Validation Loss
1.248 0.0033 1 0.7102
0.7497 0.5008 154 0.4374
0.7229 1.0016 308 0.3940
0.3772 1.4862 462 0.3962
0.3791 1.9870 616 0.3838
0.0943 2.4699 770 0.4440

Framework versions

  • Transformers 4.40.2
  • Pytorch 2.3.0+cu121
  • Datasets 2.19.1
  • Tokenizers 0.19.1
Downloads last month
50
Safetensors
Model size
70.6B params
Tensor type
BF16
·

Finetuned from

Dataset used to train shisa-ai/shisa-v1-llama3-70b

Collection including shisa-ai/shisa-v1-llama3-70b