1base_model: Qwen/Qwen3-0.6B-Base
2# Automatically upload checkpoint and final model to HF
3# hub_model_id: username/custom_model_name
4
5plugins:
6 - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
7strict: false
8
9chat_template: qwen3
10datasets:
11 - path: timarni/MNLP_M2_mcqa_dataset
12 type: alpaca
13 split: train
14
15shuffle_merged_datasets: true
16
17val_set_size: 0.1
18output_dir: ./outputs/base_test_set
19dataset_prepared_path: last_run_prepared
20
21sequence_len: 4096 #2048
22sample_packing: true # was true -> need to check if it actually learns on the samples or not (better understand te hyperparam and event. install axolotl to debug)
23eval_sample_packing: false
24pad_to_sequence_len: true
25# train_on_inputs: true # NEW
26# group_by_length: false NEW?
27
28# To be sure that no LORA is done
29adapter: null
30lora: false
31merge_lora: false
32
33wandb_project: mnlp_project
34wandb_entity: tim-arni
35wandb_watch:
36wandb_name: base_test_set
37wandb_log_model:
38
39gradient_accumulation_steps: 16 # 2
40micro_batch_size: 2 # 1
41num_epochs: 25
42optimizer: adamw_torch
43lr_scheduler: cosine
44learning_rate: 0.00005 # 0.00005
45# cosine_min_lr_ratio: 0.1
46
47warmup_ratio: 0.05
48weight_decay: 0.01
49
50bf16: auto
51tf32: true
52
53gradient_checkpointing: offload
54gradient_checkpointing_kwargs:
55 use_reentrant: false
56resume_from_checkpoint:
57logging_steps: 1
58gradient_clipping: 1.0 # or max_grad_norm?
59flash_attention: true
60
61evals_per_epoch: 4
62saves_per_epoch: 2
63save_total_limit: 25
64special_tokens:
65
