Update axolotl_config.yml

axolotl_config.yml

@@ -1,46 +1,39 @@
-base_model: mistralai/Mistral-7B-Instruct-v0.2
-model_type: MistralForCausalLM
-tokenizer_type: AutoTokenizer
-is_llama_derived_model: true
+base_model: mistralai/Mistral-7B-Instruct-v0.1
+model_type: mistral
+tokenizer_type: mistral
+chat_template: alpaca
 
 datasets:
-  - path: ./warrungu_chat_dataset.json
+  - path: /mnt/data/warrungu_chat_dataset.json
     type: alpaca
+    train_on_split: train
 
-dataset_prepared_path: ./prepared_warrungu_chat_dataset
-val_set_size: 0.05
+output_dir: ./warrungu-test-output
+logging_dir: ./logs
 
-output_dir: ./outputs/warrungu-mistral-chat
-hub_model_id: warrungu-mistral-chat-ai
-hub_strategy: every_save
+sequence_len: 512
+sample_packing: true
 
-sequence_len: 2048
-sample_packing: false
-pad_to_sequence_len: true
+adapter: lora
+lora_r: 4
+lora_alpha: 8
+lora_dropout: 0.1
+lora_target_linear: true
 
-adapter: qlora
-lora_model_dir:
-
-load_in_4bit: true
-bnb_4bit_compute_dtype: bfloat16
-bnb_4bit_quant_type: nf4
-bnb_4bit_use_double_quant: true
-
-gradient_accumulation_steps: 2
-micro_batch_size: 2
-num_epochs: 3
-optimizer: adamw_bnb_8bit
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+learning_rate: 1e-4
 lr_scheduler: cosine
-learning_rate: 2e-5
-
-save_steps: 50
-eval_steps: 10
-save_total_limit: 3
-logging_steps: 10
+optimizer: adamw_bnb_8bit
 
 train_on_inputs: false
 group_by_length: false
-bf16: auto
-fp16: false
+bf16: true
 tf32: true
-early_stopping_patience: 5
+
+save_steps: 10
+evals_per_epoch: 0
+log_freq: 1
+
+warmup_steps: 1