allura-forge/q3-30b-ft-adpt-ep2

See axolotl config

axolotl version: 0.10.0.dev0

# === Model Configuration ===
base_model: Qwen/Qwen3-30B-A3B-Base  # e.g. "mistralai/Mistral-Small-24B-Instruct-2501"
load_in_8bit: false
load_in_4bit: false

# === Training Setup ===
num_epochs: 2
micro_batch_size: 2
gradient_accumulation_steps: 1
sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

# === Hyperparameter Configuration ===
optimizer: adamw_torch_fused
# Apollo-mini configuration:
#optim_args: "proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200"
# Regular Apollo configuration:
# optim_args: 
#optim_target_modules: all_linear
learning_rate: 1e-5
lr_scheduler: rex
weight_decay: 0.01
warmup_ratio: 0.05
cosine_min_lr_ratio: 0.1

# === LoRA Configuration ===
adapter: lora
lora_r: 128
lora_alpha: 16
lora_dropout: 0.35
lora_target_modules:
lora_target_linear: true
peft_use_rslora: true

lora_mlp_kernel: true
lora_qkv_kernel: true
lora_o_kernel: true

# === Data Configuration ===
datasets:
  - path: allura-forge/fuckedup-inkmix
    type: chat_template
    split: train
    field_messages: conversations
    message_field_role: from
    message_field_content: value

dataset_prepared_path: last_run_prepared
chat_template: chatml

# === Plugins ===
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

# === Hardware Optimization ===
#51;33;32Mgradient_checkpointing: offload
#gradient_checkpointing_kwargs:
#  use_reentrant: false
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
cut_cross_entropy: true

# === Wandb Tracking ===
wandb_project: q3-30b-fuckedup-inkmix

# === Checkpointing ===
saves_per_epoch: 2
save_total_limit: 3

# === Advanced Settings ===
output_dir: ./ckpts
bf16: auto
flash_attention: true
train_on_inputs: false
group_by_length: false
logging_steps: 1
trust_remote_code: true
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_activation_checkpointing: true
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: 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: Qwen3MoeDecoderLayer
  fsdp_state_dict_type: SHARDED_STATE_DICT
  fsdp_reshard_after_forward: true
  fsdp_version: 2

ckpts

This model is a fine-tuned version of Qwen/Qwen3-30B-A3B-Base on the allura-forge/fuckedup-inkmix 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: 1e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 16
• total_eval_batch_size: 16
• 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: 38
• num_epochs: 2.0

Training results

Framework versions

• PEFT 0.15.2
• Transformers 4.51.3
• Pytorch 2.7.0+cu126
• Datasets 3.5.1
• Tokenizers 0.21.1