krinor/llama2-anon-lt

See axolotl config

axolotl version: 0.8.0.dev0

adapter: lora
base_model: neurotechnology/Lt-Llama-2-13b-instruct-hf

# mixed precision
bf16: auto

# data & splitting
dataset_processes: 32

datasets:
  # ─────────── TRAIN ───────────
  - path: .
    type: alpaca
    data_files: ["train.json"]
    message_property_mappings:
      role: role
      content: content

validation_datasets:
  # ────────── VALIDATION ──────────
  - path: .
    type: alpaca
    data_files: ["validation.json"]
    message_property_mappings:
      role: role
      content: content

# we’re using explicit splits above, so no HF split / inline splitting:
val_set_size: 0.0
shuffle_merged_datasets: false

# LoRA hyperparameters
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
  - gate_proj
  - down_proj
  - up_proj

# optimizer & schedule
optimizer: adamw_bnb_8bit
learning_rate: 2e-4
lr_scheduler: cosine
weight_decay: 0.0

# batching & accumulation
micro_batch_size: 16
gradient_accumulation_steps: 1
gradient_checkpointing: true

# training loop
num_epochs: 3
max_prompt_len: 512
sequence_len: 4096
train_on_inputs: false

# precision & quantization
load_in_8bit: true
load_in_4bit: false
qlora_sharded_model_loading: false

# resource config
use_ray: false
ray_num_workers: 1
resources_per_worker:
  GPU: 1

# output & checkpointing
output_dir: ./outputs/anon-lt-lora
save_safetensors: true
save_only_model: false
load_best_model_at_end: true
pretrain_multipack_attn: true
pretrain_multipack_buffer_size: 10000

trl:
  log_completions: false
  ref_model_sync_steps: 64
  ref_model_mixup_alpha: 0.9
  sync_ref_model: false
  use_vllm: false
  vllm_device: auto
  vllm_dtype: auto
  vllm_gpu_memory_utilization: 0.9

outputs/anon-lt-lora

This model is a fine-tuned version of neurotechnology/Lt-Llama-2-13b-instruct-hf on the None 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: 0.0002
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• 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: 3.0

Training results

Framework versions

• PEFT 0.14.0
• Transformers 4.49.0
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0