ready for eval

data/Qwen2-72B-Instruct_metrics.csv

@@ -1,2 +1,8 @@
 epoch,model,accuracy,precision,recall,f1
-0.0,Qwen/Qwen2-72B-Instruct_torch.bfloat16_4bit_lf,0.7486666666666667,0.8038991443169622,0.7486666666666667,0.7615870542896964
+0.0,Qwen/Qwen2-72B-Instruct_torch.bfloat16_4bit_lf,0.7473333333333333,0.804122252986722,0.7473333333333333,0.7607828719113865
+0.2,Qwen/Qwen2-72B-Instruct/checkpoint-35_torch.bfloat16_4bit_lf,0.7583333333333333,0.8199928526815756,0.7583333333333333,0.782751089787442
+0.4,Qwen/Qwen2-72B-Instruct/checkpoint-70_torch.bfloat16_4bit_lf,0.7366666666666667,0.8224865755517643,0.7366666666666667,0.7700627366337021
+0.6,Qwen/Qwen2-72B-Instruct/checkpoint-105_torch.bfloat16_4bit_lf,0.757,0.8253824826209251,0.757,0.784000409833628
+0.8,Qwen/Qwen2-72B-Instruct/checkpoint-140_torch.bfloat16_4bit_lf,0.7893333333333333,0.8229104753645825,0.7893333333333333,0.8033124955993173
+1.0,Qwen/Qwen2-72B-Instruct/checkpoint-175_torch.bfloat16_4bit_lf,0.7376666666666667,0.8243654864769323,0.7376666666666667,0.7699617360961548
+1.2,Qwen/Qwen2-72B-Instruct/checkpoint-210_torch.bfloat16_4bit_lf,0.763,0.8318882808702871,0.763,0.7901075708186186

scripts/eval-mgtv-qwen2_72b.sh

@@ -15,16 +15,47 @@ grep MemTotal /proc/meminfo
 
 export LOGICAL_REASONING_DATA_PATH=datasets/mgtv
 export RESIZE_TOKEN_EMBEDDINGS=true
-export START_EPOCH=0
 export USING_LLAMA_FACTORY=true
+export USING_P1_PROMPT_TEMPLATE=false
 export LOAD_IN_4BIT=true
-export MODEL_NAME=Qwen/Qwen2-72B-Instruct
 
+export START_EPOCH=7
 export MODEL_PREFIX=Qwen2-72B-Instruct_p2
+export MODEL_NAME=Qwen/$Qwen2-72B-Instruct
 
 export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX.csv
 export ADAPTER_PATH_BASE=llama-factory/saves/Qwen2-72B-Instruct
-export USING_P1_PROMPT_TEMPLATE=false
+
+echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
+python llm_toolkit/eval_logical_reasoning_all_epochs.py
+
+export START_EPOCH=0
+export MODEL_PREFIX=Llama3.1-70B-Chinese-Chat
+export MODEL_NAME=shenzhi-wang/$MODEL_PREFIX
+
+export LOGICAL_REASONING_RESULTS_PATH=data/${MODEL_PREFIX}_results.csv
+export ADAPTER_PATH_BASE=llama-factory/saves/$MODEL_PREFIX
+
+echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
+python llm_toolkit/eval_logical_reasoning_all_epochs.py
+
+
+export LOAD_IN_4BIT=false
+
+export MODEL_PREFIX=Llama3.1-8B-Chinese-Chat
+export MODEL_NAME=shenzhi-wang/$MODEL_PREFIX
+
+export LOGICAL_REASONING_RESULTS_PATH=data/${MODEL_PREFIX}_results.csv
+export ADAPTER_PATH_BASE=llama-factory/saves/$MODEL_PREFIX
+
+echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
+python llm_toolkit/eval_logical_reasoning_all_epochs.py
+
+export MODEL_PREFIX=Mistral-7B-v0.3-Chinese-Chat
+export MODEL_NAME=shenzhi-wang/$MODEL_PREFIX
+
+export LOGICAL_REASONING_RESULTS_PATH=data/${MODEL_PREFIX}_results.csv
+export ADAPTER_PATH_BASE=llama-factory/saves/$MODEL_PREFIX
 
 echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
 python llm_toolkit/eval_logical_reasoning_all_epochs.py