awilliamson/qbank-instruct

See axolotl config

axolotl version: 0.4.1

base_model: meta-llama/Meta-Llama-3-70B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: awilliamson/qbank_conversations
    type: chat_template
    chat_template: llama3
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    roles:
      system:
        - system
      user:
        - user
      assistant:
        - assistant
chat_template: llama3
adapter: qlora
lora_r: 128
lora_alpha: 32
lora_modules_to_save: [embed_tokens, lm_head]
lora_dropout: 0.05
lora_target_linear: true

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./output/llama3-70b

sequence_len: 2048
sample_packing: false
pad_to_sequence_len: true

wandb_project: llama-70b
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 2e-4

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 0
evals_per_epoch: 5
eval_table_size:
saves_per_epoch: 1
save_total_limit: 10
save_steps:
debug:
weight_decay: 0.00
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
special_tokens:
  pad_token: "<|end_of_text|>"

output/llama3-70b

This model is a fine-tuned version of meta-llama/Meta-Llama-3-70B-Instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.5806

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
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.6238	0.0769	1	1.6328
1.2354	0.2308	3	1.6006
1.1512	0.4615	6	1.6043
1.1183	0.6923	9	1.5402
1.0818	0.9231	12	1.4909
0.7404	1.1538	15	1.4745
0.6681	1.3846	18	1.5023
0.6163	1.6154	21	1.5385
0.6596	1.8462	24	1.5612
0.5081	2.0769	27	1.5699
0.5118	2.3077	30	1.5786
0.4827	2.5385	33	1.5808
0.4768	2.7692	36	1.5800
0.484	3.0	39	1.5806

Framework versions

• PEFT 0.11.1
• Transformers 4.41.1
• Pytorch 2.1.2+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1