See axolotl config

axolotl version: 0.12.2

base_model: mistralai/Mistral-7B-v0.1
# optionally might have model_type or tokenizer_type
model_type: MistralForCausalLM
# Automatically upload checkpoint and final model to HF
tokenizer_type: LlamaTokenizer
hub_model_id: anselmlong/almost-anselm

load_in_4bit: true

chat_template: chatml
datasets:
  - path: data/processed/sft_train_new.json
    type: chat_template
    field_messages: messages
    message_property_mappings:
      role: from
      content: value

    roles:
      assistant:
        - system
        - gpt
        - model
        - assistant
      user:
        - human
        - user

    # step 3
    roles_to_train: ["assistant"]
    train_on_eos: "turn"

dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: models/base_v5 

adapter: qlora
lora_model_dir:

sequence_len: 768
sample_packing: false


lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: false
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
auto_resume_from_checkpoints: true
logging_steps: 50
flash_attention: false

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_ratio: 0.1
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.0
special_tokens:

save_first_step: true  # uncomment this to validate checkpoint saving works with your config

almost-anselm

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

• Loss: 3.6523
• Memory/max Mem Active(gib): 6.62
• Memory/max Mem Allocated(gib): 6.62
• Memory/device Mem Reserved(gib): 7.07

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: 8
• total_train_batch_size: 16
• 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: 38
• training_steps: 381

Training results

Training Loss	Epoch	Step	Validation Loss	Mem Active(gib)	Mem Allocated(gib)	Mem Reserved(gib)
No log	0	0	5.0383	6.62	6.62	6.89
3.9016	0.2521	96	3.7967	6.62	6.62	7.07
3.4949	0.5041	192	3.6951	6.62	6.62	7.07
3.5013	0.7562	288	3.6523	6.62	6.62	7.07

Framework versions

• PEFT 0.17.0
• Transformers 4.55.2
• Pytorch 2.6.0+cu124
• Datasets 4.0.0
• Tokenizers 0.21.4