Mawdistical-Brew/infracelestial-lite-LoRA

See axolotl config

axolotl version: 0.16.0.dev0

# =============================================================================
# Weights and Biases logging config
# =============================================================================
wandb_project: Infracelestial
wandb_name: qlora-sft

# =============================================================================
# Model + Saving
# =============================================================================
base_model: Mawdistical/Kuwutu-7B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true
chat_template: chatml

output_dir: ./output
saves_per_epoch: 50
save_safetensors: true
save_total_limit: 5

# =============================================================================
# MIXED PRECISION
# =============================================================================
bf16: true
fp16: false
tf32: false

# =============================================================================
# MODEL LOADING — QLoRA (4-bit NF4)
# =============================================================================
load_in_8bit: false
load_in_4bit: true
strict: false

# =============================================================================
# SEQUENCE CONFIG
# =============================================================================
sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true
train_on_inputs: false
group_by_length: false

# =============================================================================
# QLoRA ADAPTER
# =============================================================================
adapter: qlora
lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
peft_use_dora: false
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

# =============================================================================
# DATASET CONFIGURATION
# =============================================================================
datasets:
  - path: data/data.jsonl
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
    roles:
      system:
        - system
      user:
        - human
      assistant:
        - gpt

shuffle_merged_datasets: true
dataset_prepared_path: ./dataset_prepared

# =============================================================================
# EVALUATION
# =============================================================================
val_set_size: 0.0

# =============================================================================
# TRAINING PARAMETERS
# =============================================================================
num_epochs: 2
micro_batch_size: 2
gradient_accumulation_steps: 4
warmup_steps: 500
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 1e-5
loraplus_lr_ratio: 8
cosine_min_lr_ratio: 0.1
weight_decay: 0.1
max_grad_norm: 1
logging_steps: 10

# =============================================================================
# MEMORY OPTIMIZATION
# =============================================================================
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
flash_attention: true

# =============================================================================
# LOSS MONITORING
# =============================================================================
early_stopping_patience:
loss_watchdog_threshold: 100.0
loss_watchdog_patience: 3

# =============================================================================
# ADDITIONAL SETTINGS
# =============================================================================
debug: false
seed: 42

deepspeed: deepspeed_configs/zero2.json

output

This model is a fine-tuned version of Mawdistical/Kuwutu-7B on the data/data.jsonl 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: 1e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT 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: 500
• training_steps: 22

Training results

Framework versions

• PEFT 0.18.1
• Transformers 5.5.3
• Pytorch 2.8.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2