base_model: mistralai/Mistral-7B-Instruct-v0.3
# optionally might have model_type or tokenizer_type
model_type: AutoModelForCausalLM
#tokenizer_type: LlamaTokenizer
# Automatically upload checkpoint and final model to HF
hub_model_id: AiAF/Mistral-7B-Instruct-v0.3-QLoRA-SFT-75567

load_in_8bit: false
load_in_4bit: true

datasets:
  - path: ./
    data_files:
      - STLD-S01E01-SFT-JSONL-fixed.jsonl
    type: chat_template
    chat_template: chatml
    field_messages: messages
    message_property_mappings:
      role: role
      content: content
    roles_to_train: ["assistant", "user", "system"]
    train_on_eos: "turn"

dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: ./outputs/qlora-out/Mistral-7B-Instruct-v0.3-QLoRA-SFT-75567
save_total_limit: 25

adapter: qlora
lora_model_dir:
lora_modules_to_save: ["embed_tokens", "lm_head"]

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: False

lora_r: 256
lora_alpha: 512
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: "LLM-SFT"
wandb_watch: "all"
wandb_name: "Mistral-7B-Instruct-v0.3-QLoRA-SFT-75567-3"
wandb_log_model: "false"
wandb_run_id: "Mistral-7B-Instruct-v0.3-QLoRA-SFT-75567-3"

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 50
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.000005

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 5
evals_per_epoch:
saves_per_epoch: 1
weight_decay: 0.0
special_tokens:
  eos_token: "<|im_end|>"

Mistral-7B-Instruct-v0.3-QLoRA-SFT-75567

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 on an unknown dataset.

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: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 8
optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 5
num_epochs: 50.0

Training results

Framework versions

PEFT 0.15.2
Transformers 4.52.4
Pytorch 2.7.0+cu126
Datasets 3.6.0
Tokenizers 0.21.1