srai86825/qwen2.5-7b-alpaca-full-lora

See axolotl config

axolotl version: 0.8.0.dev0

adapter: lora
base_model: Qwen/Qwen2.5-7B-Instruct
bf16: true  # Changed to true for full precision (use bf16 if supported by GPU, otherwise use fp16)
load_in_4bit: false  # No quantization for adapter
load_in_8bit: false

dataset_processes: 32
datasets:
- message_property_mappings:
    content: content
    role: role
  path: tatsu-lab/alpaca
  trust_remote_code: false
  type: alpaca

gradient_accumulation_steps: 4
gradient_checkpointing: true

learning_rate: 0.0002
lisa_layers_attribute: model.layers

load_best_model_at_end: false

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

loraplus_lr_embedding: 1.0e-06

lr_scheduler: cosine

max_prompt_len: 512
mean_resizing_embeddings: false
micro_batch_size: 4

num_epochs: 1.0
optimizer: adamw_torch  # Changed from 8-bit optimizer to full-precision

output_dir: ./outputs/mymodel-full-lora

pretrain_multipack_attn: true
pretrain_multipack_buffer_size: 10000

qlora_sharded_model_loading: false  # No QLoRA used now

ray_num_workers: 1
resources_per_worker:
  GPU: 1

sample_packing_bin_size: 200
sample_packing_group_size: 100000

save_only_model: false
save_safetensors: true

sequence_len: 2048

shuffle_merged_datasets: true
skip_prepare_dataset: false
strict: false

train_on_inputs: false

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

use_ray: false
val_set_size: 0.0
weight_decay: 0.0

outputs/mymodel-full-lora

This model is a fine-tuned version of Qwen/Qwen2.5-7B-Instruct on the tatsu-lab/alpaca 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: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH 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: 97
• num_epochs: 1.0

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