See axolotl config

axolotl version: 0.4.0

base_model: mistralai/Mistral-7B-v0.1
base_model_config: mistralai/Mistral-7B-v0.1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true
hub_model_id: Mistral-of-Realms-7b

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: Akila/ForgottenRealmsWikiDataset
    data_files:
      - specific_formats/FRW-J-axolotl-completion.jsonl
    type: completion
dataset_prepared_path: 
val_set_size: 0.02
output_dir: ./qlora-out

#using lora for lower cost
adapter: lora
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj

sequence_len: 512
sample_packing: false
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

#only 2 epochs because of small dataset
gradient_accumulation_steps: 3
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
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_table_max_new_tokens: 128
saves_per_epoch: 1
debug:
#default deepspeed, can use more aggresive if needed like zero2, zero3
deepspeed: 
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

Mistral-of-Realms-7b

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the Akila/ForgottenRealmsWikiDataset dataset. It achieves the following results on the evaluation set:

• Loss: 2.1762

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: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 3
• total_train_batch_size: 6
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
2.4401	0.0	1	2.5991
2.3719	0.25	2224	2.2777
2.1262	0.5	4448	2.2483
2.3942	0.75	6672	2.2234
2.3839	1.0	8896	2.2065
2.5641	1.25	11120	2.1937
2.1295	1.5	13344	2.1821
1.7813	1.75	15568	2.1773
1.9467	2.0	17792	2.1762

Framework versions

• PEFT 0.7.2.dev0
• Transformers 4.37.0
• Pytorch 2.1.2+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0