llama3 r4 p2 results

.gitattributes

@@ -98,3 +98,4 @@ results/mgtv-results_internlm_nv4090.csv filter=lfs diff=lfs merge=lfs -text
 results/glm-4-9b_lora_sft_bf16-p1.csv filter=lfs diff=lfs merge=lfs -text
 results/llama3-8b_lora_sft_bf16-p1_en.csv filter=lfs diff=lfs merge=lfs -text
 results/mgtv-results_internlm_best.csv filter=lfs diff=lfs merge=lfs -text
+results/test_b-results_r4.csv filter=lfs diff=lfs merge=lfs -text

llama-factory/config/llama3-8b_lora_sft_bf16-p1_en.yaml

@@ -1,5 +1,6 @@
 ### model
-model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+#model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+model_name_or_path: meta-llama/Meta-Llama-3.1-8B-Instruct
 
 ### method
 stage: sft
@@ -14,15 +15,15 @@ upcast_layernorm: true
 ### dataset
 dataset: alpaca_mgtv_p1_en
 template: llama3
-cutoff_len: 4096
+cutoff_len: 8192
 max_samples: 25000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/lora/sft_bf16_p1_full_en
+output_dir: saves/llama3-8b/lora/sft_bf16_p1_full_en_r2
 logging_steps: 10
-save_steps: 117
+save_steps: 35
 plot_loss: true
 # overwrite_output_dir: true
 
@@ -40,7 +41,7 @@ ddp_timeout: 180000000
 val_size: 0.1
 per_device_eval_batch_size: 1
 eval_strategy: steps
-eval_steps: 175
+eval_steps: 35
 
 report_to: wandb
-run_name: llama3_8b_p1_en # optional
+run_name: llama3_8b_p1_en_r2 # optional

llama-factory/config/llama3-8b_lora_sft_bf16-p2_en.yaml

@@ -1,5 +1,6 @@
 ### model
-model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+#model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
+model_name_or_path: meta-llama/Meta-Llama-3.1-8B-Instruct
 
 ### method
 stage: sft
@@ -14,15 +15,15 @@ upcast_layernorm: true
 ### dataset
 dataset: alpaca_mgtv_p2_en
 template: llama3
-cutoff_len: 4096
+cutoff_len: 8192
 max_samples: 25000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/llama3-8b/lora/sft_bf16_p2_full_en
+output_dir: saves/llama3-8b/lora/sft_bf16_p2_full_en_r2
 logging_steps: 10
-save_steps: 117
+save_steps: 35
 plot_loss: true
 # overwrite_output_dir: true
 
@@ -40,7 +41,7 @@ ddp_timeout: 180000000
 val_size: 0.1
 per_device_eval_batch_size: 1
 eval_strategy: steps
-eval_steps: 175
+eval_steps: 35
 
 report_to: wandb
-run_name: llama3_8b_p2_en # optional
+run_name: llama3_8b_p2_en_r2 # optional

results/test_b-results_r4.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8caa5c7f153782389b90356cd6a59bf1b5ba976d993df896b2d0bfe666d8b64
+size 23128244

scripts/eval-mgtv-llama3_8b_en.sh

@@ -24,17 +24,17 @@ export USING_LLAMA_FACTORY=true
 export USE_ENGLISH_DATASETS=true
 export USE_BF16_FOR_INFERENCE=true
 
-export MODEL_NAME=meta-llama/Meta-Llama-3-8B-Instruct
+export MODEL_NAME=meta-llama/Meta-Llama-3.1-8B-Instruct
 export MODEL_PREFIX=llama3-8b_lora_sft_bf16
 
-export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p1_en.csv
-#export ADAPTER_PATH_BASE=llama-factory/saves/llama3-8b/lora/sft_bf16_p1_full_en
+export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p1_en_r2.csv
+export ADAPTER_PATH_BASE=llama-factory/saves/llama3-8b/lora/sft_bf16_p1_full_en_r2
 export USING_P1_PROMPT_TEMPLATE=true
 echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
 python llm_toolkit/eval_logical_reasoning_all_epochs.py
 
-export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p2_en.csv
-#export ADAPTER_PATH_BASE=llama-factory/saves/llama3-8b/lora/sft_bf16_p2_full_en
+export LOGICAL_REASONING_RESULTS_PATH=results/$MODEL_PREFIX-p2_en_r2.csv
+export ADAPTER_PATH_BASE=llama-factory/saves/llama3-8b/lora/sft_bf16_p2_full_en_r2
 export USING_P1_PROMPT_TEMPLATE=false
 echo "Eval $MODEL_NAME with $ADAPTER_PATH_BASE"
 python llm_toolkit/eval_logical_reasoning_all_epochs.py

scripts/test-mgtv.sh

@@ -13,16 +13,22 @@ cat /etc/os-release
 lscpu
 grep MemTotal /proc/meminfo
 
+pip install transformers==4.41.2
+
 export LOAD_IN_4BIT=false
 export USING_LLAMA_FACTORY=true
 
-export MODEL_NAME=internlm/internlm2_5-7b-chat-1m
+#export MODEL_NAME=internlm/internlm2_5-7b-chat-1m
 # export ADAPTER_NAME_OR_PATH=inflaton-ai/InternLM_2_5-7b_LoRA-Adapter
-export ADAPTER_NAME_OR_PATH=llama-factory/saves/internlm2_5_7b/lora/sft_bf16_p2_full_r3/checkpoint-140
+#export ADAPTER_NAME_OR_PATH=llama-factory/saves/internlm2_5_7b/lora/sft_bf16_p2_full_r3/checkpoint-140
+
+export MODEL_NAME=shenzhi-wang/Llama3-8B-Chinese-Chat
+export ADAPTER_NAME_OR_PATH=llama-factory/saves/llama3-8b/lora/sft_bf16_p1_full_r4/checkpoint-140
+
 export LOGICAL_REASONING_DATA_PATH=datasets/mgtv
 
 export TEST_DATA=test_b
-export LOGICAL_REASONING_RESULTS_PATH=results/$TEST_DATA-results_r3.csv
+export LOGICAL_REASONING_RESULTS_PATH=results/$TEST_DATA-results_r4.csv
 
 echo "Eval $MODEL_NAME with $ADAPTER_NAME_OR_PATH"
 python llm_toolkit/eval_logical_reasoning.py

scripts/tune-mgtv-llama3_8b_en.sh

@@ -18,9 +18,11 @@ grep MemTotal /proc/meminfo
 #pip install transformers==4.41.2
 #pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
 
+pip install -U transformers
+
 export LOGICAL_REASONING_DATA_PATH=datasets/mgtv
 
-export MODEL_NAME=meta-llama/Meta-Llama-3-8B-Instruct
+export MODEL_NAME=meta-llama/Meta-Llama-3.1-8B-Instruct
 
 export MODEL_PREFIX=llama3-8b_lora_sft_bf16
 

scripts/tune-mgtv.sh

@@ -1 +1 @@
-tune-mgtv-llama3_8b.sh
+tune-mgtv-llama3_8b_en.sh