1base_model: timarni/qwen3_dpo
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_M3_mcqa_dataset # timarni/MNLP_intstruction_tuning
12 name: stem_instruction_tuning_balanced_mini
13 type: alpaca
14 split: train
15
16shuffle_merged_datasets: true
17
18val_set_size: 0.1
19output_dir: ./outputs/dpo_it_bal
20dataset_prepared_path: last_run_prepared
21
22sequence_len: 2048 #2048
23sample_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)
24eval_sample_packing: true
25pad_to_sequence_len: true
26train_on_inputs: false # NEW
27# group_by_length: false NEW?
28
29# To be sure that no LORA is done
30adapter: null
31lora: false
32merge_lora: false
33
34wandb_project: mnlp_project
35wandb_entity: tim-arni
36wandb_watch:
37wandb_name: wiki_it_bal
38wandb_log_model:
39
40gradient_accumulation_steps: 2
41micro_batch_size: 1 # 2
42num_epochs: 6
43optimizer: adamw_torch
44lr_scheduler: cosine
45learning_rate: 5e-6 # 0.00005
46cosine_min_lr_ratio: 0.1
47
48warmup_ratio: 0.05
49weight_decay: 0.01
50
51bf16: auto
52tf32: true
53
54gradient_checkpointing: offload
55gradient_checkpointing_kwargs:
56 use_reentrant: false
57resume_from_checkpoint:
58logging_steps: 1
59gradient_clipping: 1.0 # or max_grad_norm?
60flash_attention: true
61
62evals_per_epoch: 4
63saves_per_epoch: 2
64save_total_limit: 25
65special_tokens:
66
