See axolotl config

axolotl version: 0.4.0

base_model: BlouseJury/Mistral-7B-Discord-0.1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: Intel/orca_dpo_pairs
    type:
      system_prompt: ""
      field_system: system
      field_instruction: question
      field_output: rejected
      field_output: chosen
      format: "[INST] {instruction} [/INST]"
      no_input_format: "[INST] {instruction} [/INST]"
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./out

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.000005

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_table_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

BlouseJury/Mistral-7B-Discord-0.1-DPO

This model is a fine-tuned version of BlouseJury/Mistral-7B-Discord-0.1 on the Intel/orca_dpo_pairs dataset. It achieves the following results on the evaluation set:

• Loss: 0.7923

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• 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
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.1157	0.01	1	1.1924
1.0146	0.26	19	0.8381
0.9004	0.51	38	0.8015
0.8425	0.77	57	0.7923

Framework versions

• Transformers 4.37.0
• Pytorch 2.1.2+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	62.29
AI2 Reasoning Challenge (25-Shot)	63.23
HellaSwag (10-Shot)	83.27
MMLU (5-Shot)	62.62
TruthfulQA (0-shot)	55.28
Winogrande (5-shot)	78.93
GSM8k (5-shot)	30.40