L3.3-70B-PippaMaid-1.0-NVFP4

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. --> <div align="center"> <b style="font-size: 40px;">L3.3-70B-PippaMaid-1.0</b> </div> <img src="https://www.gofigstudios.com/wp-content/uploads/2025/12/pippamaid.png" alt="L3.3-70B-PippaMaid-1.0" style="width: 70%; min-width: 640px; display: block; margin: auto;">

---

axolotl version: 0.10.0

base_model: huihui-ai/Llama-3.3-70B-Instruct-abliterated
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

# 4-bit quantization for QLoRA
load_in_8bit: false
load_in_4bit: true
bnb_4bit_compute_dtype: bfloat16
bnb_4bit_use_double_quant: true
bnb_4bit_quant_type: nf4

datasets:
  - path: Shifusen/LumimaidQAT
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
dataset_prepared_path: /root/last_run_prepared
val_set_size: 0.02
output_dir: ./outputs/L3.3-70B-PippaMaid-1.0

adapter: qlora
lora_model_dir:

# 8x RTX PRO 6000 Blackwell (96GB) - increased context for larger VRAM
sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

# Higher rank for 1.78M sample dataset
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true

wandb_project: L33-70B-PippaMaid
wandb_entity: gofigstudios-gofig-studios
wandb_watch: "false"
wandb_name: pippamaid-qlora-r64-8k-8gpu
wandb_log_model: "false"

# 8x RTX PRO 6000: batch 1 per GPU, 8 accum = effective batch 64
gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 10
flash_attention: true

warmup_ratio: 0.03
evals_per_epoch: 4
saves_per_epoch: 2
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: false
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
  fsdp_state_dict_type: SHARDED_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
special_tokens:
  pad_token: <|end_of_text|>

</details><br>

outputs/L3.3-70B-PippaMaid-1.0

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

• Loss: 0.8259

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: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• 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: 36
• training_steps: 1229

Training results

Framework versions

• PEFT 0.15.2
• Transformers 4.52.3
• Pytorch 2.9.1+cu128
• Datasets 3.6.0
• Tokenizers 0.21.4