Update kto-4b.yml

kto-4b.yml

@@ -1,12 +1,24 @@
-base_model: NewEden/Hamanasu-4B-R2
+base_model: ./model
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_layer_norm: true
+liger_glu_activation: true
+liger_fused_linear_cross_entropy: false
+cut_cross_entropy: true
+
+
 load_in_8bit: false
 load_in_4bit: false
 strict: false
 
-hub_model_id: NewEden/KTO-4B
+hub_model_id: NewEden/Hamanasu-KTO-2
 hub_strategy: "all_checkpoints"
 push_dataset_to_hub:
 hf_use_auth_token: true
@@ -14,17 +26,19 @@ hf_use_auth_token: true
 chat_template: chatml
 
 rl: kto
-rl_beta: 0.2
-kto_desirable_weight: 0.2
+kto_undesirable_weight: 1.0
 
 datasets:
-  - path: NewEden/KTO-IF-Dans
+  - path: NewEden/Opus-accepted-hermes-rejected-shuffled
+    split: train
+    type: chatml.argilla
+  - path: NewEden/KTO-Instruct-Mix
     split: train
     type: chatml.argilla
 dataset_prepared_path: last_run_prepared
 
 shuffle_merged_datasets: true
-val_set_size: 0.0
+val_set_size: 0.01
 output_dir: ./outputs/out
 
 sequence_len: 8192
@@ -35,22 +49,22 @@ pad_to_sequence_len: false
 wandb_project: tavbussy
 wandb_entity:
 wandb_watch:
-wandb_name: kto-1
+wandb_name: kto-chat-2
 wandb_log_model:
 
-gradient_accumulation_steps: 16
+gradient_accumulation_steps: 4
 micro_batch_size: 2
 num_epochs: 1
-optimizer: paged_adamw_8bit
-learning_rate: 1e-6
+optimizer: adamw_bnb_8bit
+learning_rate: 5e-7
 max_grad_norm: 0.001
-lr_scheduler: cosine
+lr_scheduler: constant_with_warmup
 weight_decay: 0.02
 
 
 lora_r: 64
 lora_alpha: 32
-lora_dropout: 0.0
+lora_dropout: 0.1
 lora_target_linear: true
 lora_fan_in_fan_out:
 lora_target_modules: