new lr

examples/yayi2-30b/qlora.yml

@@ -3,7 +3,7 @@ model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 is_mistral_derived_model: false
 trust_remote_code: true
-
+model_revision: refs/pr/5
 
 load_in_8bit: false
 load_in_4bit: true
@@ -19,7 +19,7 @@ output_dir: ./qlora-out
 adapter: qlora
 lora_model_dir:
 
-sequence_len: 8192
+sequence_len: 2048 # Fits in 40gb VRAM. Can easily do 4096 in A100 80 or a A6000
 sample_packing: false
 pad_to_sequence_len: true
 
@@ -37,7 +37,7 @@ lora_target_modules:
   - k_proj
   - o_proj
 
-wandb_project:
+wandb_project: yayi2
 wandb_entity:
 wandb_watch:
 wandb_name:
@@ -48,7 +48,7 @@ micro_batch_size: 1
 num_epochs: 1
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
-learning_rate: 0.0002
+learning_rate: 0.00002
 
 train_on_inputs: false
 group_by_length: false