base_model: cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer

load_in_8bit: true
load_in_4bit: false
strict: false

datasets:
  - path: ./julia/data.jsonl
    type: sharegpt
    conversation: chatml
dataset_prepared_path: ./julia/prepared_data
chat_template: chatml
val_set_size: 0.05
output_dir: ./julia/lora-out
hub_model_id: animmina/cheater-7b
hub_strategy: every_save
hf_use_auth_token: true

sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

adapter: lora
lora_model_dir:
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project: cheater-7b
wandb_entity:
wandb_watch:
wandb_name: v02
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00003

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

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

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "<|im_end|>"
  unk_token: "<unk>"

cheater-7b

This model is a fine-tuned version of cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser on 11 test cases from the Julia LLM Leaderboard. It achieves the following results on the evaluation set:

Loss: 0.5741

Model description

Simple LORA adapter (rank: 8).

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: 3e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 10
num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
0.554	0.04	1	0.6521
0.4152	0.26	7	0.6499
0.3984	0.52	14	0.6283
0.4133	0.78	21	0.6140
0.3772	1.04	28	0.5951
0.3855	1.22	35	0.5869
0.4077	1.48	42	0.5840
0.3104	1.74	49	0.5793
0.3345	2.0	56	0.5776
0.3207	2.19	63	0.5761
0.3679	2.44	70	0.5784
0.3593	2.7	77	0.5781
0.2391	2.96	84	0.5761
0.3329	3.15	91	0.5743
0.2636	3.41	98	0.5744
0.3114	3.67	105	0.5741

Framework versions

PEFT 0.9.0
Transformers 4.38.2
Pytorch 2.1.2+cu121
Datasets 2.18.0
Tokenizers 0.15.0

svilupp/cheater-7b