L3.3 70B Euryale V2.3

axolotl版本: 0.5.2

base_model: meta-llama/Llama-3.3-70B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false
sequence_len: 16384
bf16: auto
fp16:
tf32: false
flash_attention: true

adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 16
lora_dropout: 0.1
lora_target_linear: true
lora_fan_in_fan_out:
peft_use_rslora: true

# 数据
dataset_prepared_path: last_run_prepared
datasets:
  - path: datasets/amoral-full-sys-prompt.json # 未对齐数据 - 从原始数据清理而来，拆分到单独文件
    type: customllama3
  - path: datasets/mimi-superfix-RP-filtered-fixed.json # 角色扮演/创意指令数据
    type: customllama3
  - path: datasets/hespera-smartshuffle.json # Hesperus-v2-Instruct数据
    type: customllama3
warmup_steps: 15

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

# 迭代次数
num_epochs: 1

# 批处理
gradient_accumulation_steps: 4
micro_batch_size: 1
gradient_checkpointing: "unsloth"

# 优化器
optimizer: paged_ademamix_8bit
lr_scheduler: cosine
learning_rate: 0.000004
weight_decay: 0.1
max_grad_norm: 25.0

# 迭代次数
num_epochs: 1

# 其他
deepspeed: ./deepspeed_configs/zero3_bf16.json