base_model: mistralai/Mistral-7B-v0.1
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
tokenizer_use_fast: false
resize_token_embeddings_to_32x: true

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: chatml
datasets:
  - path: skymizer/open-orca-conversations
    type: chat_template
    field_messages: messages

hf_use_auth_token: true
dataset_prepared_path: pretokenized/open-orca
output_dir: ./outputs/out

sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

val_set_size: 0.005
eval_sample_packing: false
# eval_causal_lm_metrics: ["perplexity"]

wandb_project: "axolotl_mistral_sft"
wandb_entity:
wandb_watch:
wandb_name: "mistral-7B-v0.1-dense-sft-open-orca-dry-run"
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 8
max_steps: 2000
optimizer: adamw_torch
lr_scheduler: cosine
cosine_min_lr_ratio: 0.1
learning_rate: 0.000005 
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.95
adam_eps: 0.000001
max_grad_norm: 1.0

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

hub_model_id: "skymizer/mistral-7b-v0.1-sft-open-orca-dry-run"
save_strategy: "steps"
save_steps: 1000

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_ratio: 0.03
eval_steps: 500
# evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"
  pad_token: "<|im_end|>"
tokens:
  - "<|im_start|>"

seed: 42

mistral-7b-v0.1-sft-open-orca-dry-run

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the None dataset. It achieves the following results on the evaluation set:

Loss: 2.0993

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: 5e-06
train_batch_size: 8
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 4
total_train_batch_size: 128
total_eval_batch_size: 32
optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 60
training_steps: 2000

Training results

Training Loss	Epoch	Step	Validation Loss
2.18	0.0002	1	4.8865
0.5965	0.0820	500	1.9648
0.5088	0.1639	1000	2.0442
0.4913	0.2459	1500	2.0902
0.4941	0.3279	2000	2.0993

Framework versions

Transformers 4.46.3
Pytorch 2.5.1+cu124
Datasets 3.1.0
Tokenizers 0.20.3