See axolotl config

axolotl version: 0.11.0.dev0

adapter: lora
base_model: MuXodious/Mistral-Nemo-Instruct-2407-absolute-heresy
bf16: true
datasets:
- ds_type: json
  path: DarkArtsForge/Poe_v1
  type: alpaca
flash_attention: true
fp16: false
gradient_accumulation_steps: 4
gradient_checkpointing: true
hub_always_push: true
hub_model_id: Naphula-Archives/MN-Raven-12B-v0o-Instruct-LoRA
hub_private_repo: true
hub_strategy: checkpoint
is_mistral_derived_model: true
learning_rate: 1.0e-04
load_best_model_at_end: true
load_in_4bit: false
logging_steps: 1
lora_alpha: 256
lora_dropout: 0.05
lora_r: 128
lora_target_linear: false
lora_target_modules:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj
lr_scheduler: cosine
max_grad_norm: 0.3
warmup_ratio: 0.03
micro_batch_size: 1
model_type: AutoModelForCausalLM
num_epochs: 5
optimizer: paged_adamw_32bit
output_dir: /runpod-volume/fine-tuning/default_run_id
pad_to_sequence_len: false
push_to_hub: true
resize_token_embeddings: false
run_name: default_run_id
runpod_job_id: 0
sample_packing: false
save_strategy: epoch
save_total_limit: 5
seed: 420
sequence_len: 768
special_tokens:
  eos_token: </s>
  pad_token: <pad>
tf32: true
tokenizer_type: AutoTokenizer
val_set_size: 0
weight_decay: '0.0'

MN-Raven-12B-v0o-Instruct-LoRA

This model is a fine-tuned version of MuXodious/Mistral-Nemo-Instruct-2407-absolute-heresy on the DarkArtsForge/Poe_v1 dataset.

Model description

    {
      "epoch": 3.0,
      "grad_norm": 1.7338424921035767,
      "learning_rate": 3.726644398793857e-05,
      "loss": 0.1262,
      "step": 75
    }

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: 1.0e-04
• train_batch_size: 1
• eval_batch_size: 1
• seed: 420
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 15
• training_steps: 500

Training results

Framework versions

• PEFT 0.15.2
• Transformers 4.52.4
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.1