See axolotl config

axolotl version: 0.8.1

base_model: huihui-ai/Llama-3.3-70B-Instruct-abliterated-finetuned
load_in_8bit: false
load_in_4bit: true
adapter: qlora
wandb_name: 5_70B_axolotl_ft
output_dir: ./outputs/out/5_70B_axolotl_ft
hub_model_id: cgifbribcgfbi/Llama-3.3-70B-Instruct-abliterated-finetuned-chemistry-5000-e4
hub_strategy: every_save
resume_from_checkpoint: ./outputs/out/5_70B_axolotl_ft/checkpoint-72

tokenizer_type: AutoTokenizer
push_dataset_to_hub:
strict: false

datasets:
  - path: no_splits_5000_benign.jsonl
    type: chat_template
    split: train
    roles_to_train: ["assistant"]
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
# test_datasets:
#   - path: 5000_benign_val.json
#     type: chat_template
#     split: train
save_safetensors: true

sequence_len: 2800
sample_packing: true
pad_to_sequence_len: true

lora_r: 128
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true

wandb_mode:
wandb_project: finetune-chem
wandb_entity: gpoisjgqetpadsfke
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 4
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 0.00002

train_on_inputs: false
group_by_length: true
bf16: true
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true
logging_steps: 1
flash_attention: true

warmup_steps: 10
evals_per_epoch: 3
saves_per_epoch: 1
weight_decay: 0.01
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: false
  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: <|finetune_right_pad_id|>

Llama-3.3-70B-Instruct-abliterated-finetuned-chemistry-5000-e4

This model is a fine-tuned version of huihui-ai/Llama-3.3-70B-Instruct-abliterated-finetuned on the no_splits_5000_benign.jsonl dataset. It achieves the following results on the evaluation set:

• Loss: 0.3188

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: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 32
• total_eval_batch_size: 32
• 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: 10
• num_epochs: 4.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.5853	0.0139	1	0.5987
0.5321	0.3333	24	0.5129
0.432	0.6667	48	0.4115
0.3968	1.0	72	0.3711
0.3559	1.3333	96	0.3513
0.3273	1.6667	120	0.3392
0.3329	2.0	144	0.3316
0.3446	2.3333	168	0.3260
0.3229	2.6667	192	0.3228
0.3194	3.0	216	0.3204
0.3179	3.3333	240	0.3194
0.3381	3.6667	264	0.3189
0.3323	4.0	288	0.3188

Framework versions

• PEFT 0.15.1
• Transformers 4.51.0
• Pytorch 2.5.1+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1