llama3 p2 epoch 1 results

llama-factory/config/llama3-8b_lora_sft_bf16-p1.yaml

@@ -25,7 +25,7 @@ output_dir: saves/llama3-8b/lora/sft_bf16_p1_full
 logging_steps: 10
 save_steps: 175
 plot_loss: true
-#overwrite_output_dir: true
+overwrite_output_dir: true
 
 ### train
 per_device_train_batch_size: 16

scripts/eval-mgtv-llama3_8b.sh

@@ -17,7 +17,6 @@ grep MemTotal /proc/meminfo
 #cd ../LLaMA-Factory && pip install -e .[torch,bitsandbytes]
 
 export LOGICAL_REASONING_DATA_PATH=datasets/mgtv
-export START_EPOCH=1
 export RESIZE_TOKEN_EMBEDDINGS=true
 
 #export MODEL_NAME=FlagAlpha/Llama3-Chinese-8B-Instruct
@@ -27,11 +26,13 @@ export MODEL_PREFIX=llama3-8b_lora_sft_bf16
 export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p1.csv
 export ADAPTER_PATH_BASE=llama-factory/saves/llama3-8b/lora/sft_bf16_p1_full
 export USING_P1_PROMPT_TEMPLATE=true
+export START_EPOCH=3
 echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
-python llm_toolkit/eval_logical_reasoning_all_epochs.py
+#python llm_toolkit/eval_logical_reasoning_all_epochs.py
 
 export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p2.csv
 export ADAPTER_PATH_BASE=llama-factory/saves/llama3-8b/lora/sft_bf16_p2_full
 export USING_P1_PROMPT_TEMPLATE=false
+export START_EPOCH=1
 echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
 python llm_toolkit/eval_logical_reasoning_all_epochs.py