Update mixtral-zhtw.yml

mixtral-zhtw.yml

@@ -1,6 +1,7 @@
 base_model: mistralai/Mixtral-8x7B-Instruct-v0.1
-model_type: AutoModelForCausalLM
+model_type: MixtralForCausalLM
 tokenizer_type: LlamaTokenizer
+is_mistral_derived_model: false
 trust_remote_code: true
 
 load_in_8bit: false
@@ -10,24 +11,15 @@ strict: false
 datasets:
   - path: yentinglin/v1
     type: sharegpt
-    conversation: mistral
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.0
-output_dir: ./qlora-out-3e
-
-## You can optionally freeze the entire model and unfreeze a subset of parameters
-unfrozen_parameters:
-#  - lm_head.*
-#  - model.embed_tokens.*
-#  - model.layers.2[0-9]+.block_sparse_moe.gate.*
-#  - model.layers.2[0-9]+.block_sparse_moe.experts.*
-#  - model.layers.3[0-9]+.block_sparse_moe.gate.*
-#  - model.layers.3[0-9]+.block_sparse_moe.experts.*
+    conversation: chatml
+dataset_prepared_path:
+val_set_size: 0.005
+output_dir: ./Taiwan-LLM-MoE-pilot-v0.2
 
 adapter: qlora
 lora_model_dir:
 
-sequence_len: 4096
+sequence_len: 16384
 sample_packing: true
 pad_to_sequence_len: true
 
@@ -37,17 +29,12 @@ model_config:
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
+lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out:
-#lora_target_modules:
-#  - gate
-#  - q_proj
-#  - k_proj
-#  - v_proj
-#  - o_proj
-#  - w1
-#  - w2
-#  - w3
+lora_modules_to_save:
+  - embed_tokens
+  - lm_head
 
 hub_model_id: yentinglin/Taiwan-LLM-MoE-chat-alpha-3e
 hub_strategy: end
@@ -57,10 +44,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
-gradient_accumulation_steps: 4
-micro_batch_size: 1
+gradient_accumulation_steps: 2
+micro_batch_size: 3
 num_epochs: 3
-optimizer: adamw_bnb_8bit
+optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
@@ -78,9 +65,6 @@ logging_steps: 1
 xformers_attention:
 flash_attention: true
 
-loss_watchdog_threshold: 5.0
-loss_watchdog_patience: 3
-
 warmup_steps: 10
 evals_per_epoch: 4
 eval_table_size:
@@ -92,6 +76,10 @@ weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
+  eos_token: "<|im_end|>"
+tokens:
+  - "<|im_start|>"
+trust_remote_code: true
 
 ddp_timeout: 8640000000
 dataset_processes: 16