See axolotl config

axolotl version: 0.4.0

base_model: ./models/scb10x_typhoon-7b
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false


datasets:
  - path: ./work/thai_food.json
    type: completion

dataset_prepared_path: ./work/last_run_prepared
val_set_size: 0.1
output_dir: ./work/out


adapter: qlora
lora_model_dir:

sequence_len: 4096
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: true

gpu_memory_limit: 20

lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:


wandb_project: typhoon-7b
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0004

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

gradient_checkpointing: true
early_stopping_patience: 3
resume_from_checkpoint: false
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

# loss_watchdog_threshold: 5.0
# loss_watchdog_patience: 3

warmup_ratio: 0.01
# evals_per_epoch: 10
eval_steps: 2
eval_table_size:
eval_table_max_new_tokens: 128
# saves_per_epoch: 10
save_steps: 2
save_total_limit: 20
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:

ping98k/typhoon-thai-food-lora

This model was trained from thai_food dataset but re-order header to เครื่องปรุง -> วิธีทำ -> ชื่ออาหาร. It achieves the following results on the evaluation set:

• Loss: 1.9505

Model description

fill ingredients then model will create new menu.

prompt

you can let model fill more ingredients by remove ## วิธีทำ from prompt

input

## เครื่องปรุง
- ไข่เป็ด
- ใบเตย

or

## เครื่องปรุง
- ไข่เป็ด
- ใบเตย

## วิธีทำ

output

ปอกไข่ แช่น้ำใบเตยให้ทั่ว แล้วใส่ชามแช่ไว้ประมาณ 15 นาที ยกขึ้นล้างน้ำเย็นจัด (อย่าใช้น้ำแข็ง) จึงแกะสลัก

## ชื่ออาหาร
ไข่เป็ดตุ๋นใบเตย

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: 0.0004
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
2.8268	0.13	2	2.4822
2.4085	0.25	4	2.2715
2.2752	0.38	6	2.1985
2.4104	0.51	8	2.1000
2.0149	0.63	10	2.0255
2.1234	0.76	12	1.9926
2.2013	0.89	14	1.9894
1.8355	1.02	16	1.9684
1.4604	1.14	18	1.9610
1.6539	1.27	20	1.9517
1.5531	1.4	22	1.9414
1.4649	1.52	24	1.9230
1.464	1.65	26	1.9214
1.3731	1.78	28	1.9116
1.4451	1.9	30	1.8922
1.3635	2.03	32	1.8885
1.1453	2.16	34	1.9034
1.0397	2.29	36	1.9281
0.9735	2.41	38	1.9505

Framework versions

• PEFT 0.7.1
• Transformers 4.37.0
• Pytorch 2.0.1+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0