Views
No views yet
0.10.0.dev01# === Model Configuration ===
2base_model: Qwen/Qwen3-8B-Base
3load_in_8bit: false
4load_in_4bit: false
5
6# === Training Setup ===
7num_epochs: 2
8micro_batch_size: 32
9gradient_accumulation_steps: 1
10sequence_len: 8192
11sample_packing: true
12pad_to_sequence_len: true
13
14# === Hyperparameter Configuration ===
15optimizer: apollo_adamw_layerwise
16# Apollo-mini configuration:
17optim_args: "proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200"
18# Regular Apollo configuration:
19# optim_args:
20optim_target_modules: all_linear
21learning_rate: 2e-5
22lr_scheduler: rex
23weight_decay: 0.01
24warmup_ratio: 0
25
26# === Data Configuration ===
27datasets:
28 - path: allura-org/inkmix-v3.0
29 type: chat_template
30 split: train
31 field_messages: conversations
32 message_field_role: from
33 message_field_content: value
34
35dataset_prepared_path: last_run_prepared
36chat_template: chatml
37
38# === Plugins ===
39plugins:
40 - axolotl.integrations.liger.LigerPlugin
41 - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
42
43# === Hardware Optimization ===
44gradient_checkpointing: unsloth
45gradient_checkpointing_kwargs:
46 use_reentrant: false
47liger_rope: true
48liger_rms_norm: true
49liger_glu_activation: true
50cut_cross_entropy: true
51
52# === Wandb Tracking ===
53wandb_project: qwen3-8b-inkmix-v3
54
55# === Checkpointing ===
56saves_per_epoch: 2
57save_total_limit: 3
58
59# === Advanced Settings ===
60output_dir: /ephemeral/ckpts
61bf16: auto
62flash_attention: true
63train_on_inputs: false
64group_by_length: false
65logging_steps: 1
66trust_remote_code: true
67