adds updated model paths, adds eval to sft scripts

recipes/zephyr-7b/dpo/config_full.yaml

@@ -1,5 +1,5 @@
 # Model arguments
-model_name_or_path: lewtun/zephyr-7b-sft
+model_name_or_path: alignment-handbook/zephyr-7b-sft-full
 
 # Data training arguments
 # For definitions, see: src/h4/training/config.py

recipes/zephyr-7b/dpo/config_lora.yaml

@@ -1,6 +1,5 @@
 # Model arguments
-model_name_or_path: HuggingFaceH4/mistral-7b-ift
-model_revision: v14.0
+model_name_or_path: alignment-handbook/zephyr-7b-sft-lora
 torch_dtype: auto
 
 # LORA

recipes/zephyr-7b/sft/config_full.yaml

@@ -14,6 +14,7 @@ preprocessing_num_workers: 12
 
 # SFT trainer config
 bf16: true
+do_eval: true
 evaluation_strategy: epoch
 gradient_accumulation_steps: 2
 gradient_checkpointing: true

recipes/zephyr-7b/sft/config_lora.yaml

@@ -24,6 +24,7 @@ preprocessing_num_workers: 12
 
 # SFT trainer config
 bf16: true
+do_eval: true
 evaluation_strategy: epoch
 gradient_accumulation_steps: 128
 ddp_find_unused_parameters: true