al9sr31/KoSOLAR-10.7B-test-v0.2

See axolotl config

axolotl version: 0.4.0

base_model: TheBloke/SOLAR-10.7B-Instruct-v1.0-uncensored-GPTQ
is_llama_derived_model: false
gptq: true
gptq_disable_exllama: true
model_type: AutoModelForCausalLM
tokenizer_type: LlamaTokenizer
tokenizer_use_fast: true
tokenizer_legacy: true
load_in_8bit: false
load_in_4bit: false
strict: false
push_dataset_to_hub:
hf_use_auth_token: true
datasets:
  - path:  datasets_cleansinng/datasets/helper_selector_1280_0305_v01.jsonl #Path to json dataset file in huggingface
    #for type,conversation arguments read axolotl readme and pick what is suited for your project, I wanted a chatbot and put sharegpt and chatml
    type: 
      system_prompt: "Instruction에 따라 적절하게 Input 데이터를 활용하여 Output 답변을 하세요. 너는 사용자 질문(Instruction)에 실시간으로 API 호출을 위한 Json 형식의 구조화된 결과를 생성하는 인공지능이야."
      format: "[INST]### Instruction:\n{instruction}\n\n### Input:{input}\n\n[/INST]### Output: "
      no_input_format: "[INST]### Instruction:\n{instruction}\n\n[/INST]### Output: "
      field_instruction: Instruction 
      field_input: Input
      field_output: Output
dataset_prepared_path:
val_set_size: 0.05
adapter: lora
lora_model_dir:
sequence_len: 4096
sample_packing:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - k_proj
  - o_proj
  - q_proj
  - v_proj
lora_target_linear:
lora_fan_in_fan_out:
wandb_project:
wandb_watch:
wandb_name:
wandb_log_model:
output_dir: ./output_solor/exp_16
gradient_accumulation_steps: 8
micro_batch_size: 8
num_epochs: 5
optimizer: adamw_torch
adam_beta2: 0.95
adam_eps: 0.00001
max_grad_norm: 1.0
torchdistx_path:
lr_scheduler: cosine
lr_quadratic_warmup: true
learning_rate: 0.0005
train_on_inputs: false
group_by_length: false
bf16: false
fp16: false
float16: true
tf32: true
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention:
sdp_attention:
flash_optimum:
warmup_steps: 100
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero1.json
weight_decay: 0.1
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

output_solor/exp_16

This model is a fine-tuned version of TheBloke/SOLAR-10.7B-Instruct-v1.0-uncensored-GPTQ on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.2015

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: 0.0005
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss
1.3493	0.05	1	1.2795
1.2483	0.26	5	1.2769
1.2275	0.53	10	1.2099
1.0529	0.79	15	1.0724
0.8642	1.05	20	0.9709
0.8477	1.32	25	0.8245
0.7207	1.58	30	0.6994
0.4656	1.84	35	0.5878
0.4949	2.11	40	0.4970
0.3497	2.37	45	0.4221
0.3288	2.63	50	0.3672
0.3011	2.89	55	0.3250
0.2648	3.16	60	0.2900
0.3084	3.42	65	0.2591
0.2696	3.68	70	0.2459
0.2197	3.95	75	0.2286
0.1905	4.21	80	0.2111
0.1815	4.47	85	0.2084
0.2164	4.74	90	0.2128
0.1412	5.0	95	0.2015

Framework versions

• PEFT 0.9.1.dev0
• Transformers 4.38.0.dev0
• Pytorch 2.0.1+cu117
• Datasets 2.18.0
• Tokenizers 0.15.0