你的update 34B

Author: JulietLJY

Readme.md

@@ -7,7 +7,7 @@
 [![PWC](https://img.shields.io/endpoint?url=https%3A%2F%2Fpaperswithcode.com%2Fbadge%2Fmotcoder-elevating-large-language-models-with%2Fcode-generation-on-apps%3Fmetric%3DIntroductory%2520Pass%25401)](https://paperswithcode.com/sota/code-generation-on-apps?metric=Introductory%20Pass%401/motcoder-elevating-large-language-models-with) 
 [![PWC](https://img.shields.io/endpoint?url=https%3A%2F%2Fpaperswithcode.com%2Fbadge%2Fmotcoder-elevating-large-language-models-with%2Fcode-generation-on-codecontests%3Fmetric%3DTest%2520Set%2520pass%25401)](https://paperswithcode.com/sota/code-generation-on-codecontests?metric=Test%20Set%20pass%401)
 
-Large Language Models (LLMs) have showcased impressive capabilities in handling straightforward programming tasks. However, their performance tends to falter when confronted with more challenging programming problems. We observe that conventional models often generate solutions as monolithic code blocks, restricting their effectiveness in tackling intricate questions. To overcome this limitation, we present Module-of-Thought Coder (MoTCoder). We introduce a framework for MoT instruction tuning, designed to promote the decomposition of tasks into logical sub-tasks and sub-modules. Our investigations reveal that, through the cultivation and utilization of sub-modules, MoTCoder significantly improves both the modularity and correctness of the generated solutions, leading to substantial pass@1 improvements of 2.4% on APPS and 4.5% on CodeContests. MoTCoder also achieved significant improvements in self-correction capabilities, surpassing the current SOTA by 3.3%. Additionally, we provide an analysis of between problem complexity and optimal module decomposition and evaluate the maintainability index, confirming that the code generated by MoTCoder is easier to understand and modify, which can be beneficial for long-term code maintenance and evolution. Our codes are available at https://github.com/dvlab-research/MoTCoder.
+Large Language Models (LLMs) have showcased impressive capabilities in handling straightforward programming tasks. However, their performance tends to falter when confronted with more challenging programming problems. We observe that conventional models often generate solutions as monolithic code blocks, restricting their effectiveness in tackling intricate questions. To overcome this limitation, we present Module-of-Thought Coder (MoTCoder). We introduce a framework for MoT instruction tuning, designed to promote the decomposition of tasks into logical sub-tasks and sub-modules. Our investigations reveal that, through the cultivation and utilization of sub-modules, MoTCoder significantly improves both the modularity and correctness of the generated solutions, leading to substantial pass@1 improvements of 5.9% on APPS and 5.8% on CodeContests. MoTCoder also achieved significant improvements in self-correction capabilities, surpassing the current SOTA by 3.3%. Additionally, we provide an analysis of between problem complexity and optimal module decomposition and evaluate the maintainability index, confirming that the code generated by MoTCoder is easier to understand and modify, which can be beneficial for long-term code maintenance and evolution. Our codes are available at https://github.com/dvlab-research/MoTCoder.
 
 <div style="text-align: center;">
     <img src="./imgs/impression.png" alt="impression" />
@@ -56,17 +56,17 @@ Download [MoTCoder-7B-v1.5](https://huggingface.co/JingyaoLi/MoTCoder-7B-v1.5) a
 ```
 2. Run the following script to prepare the test data:
 ```bash
-python /mnt/nas-alinlp/ljy/MoTCoder/MoTCoder/eval/dataset/data.py
+python eval/dataset/data.py
 ```
 
 ### Run Evaluation
 Execute the scripts below to perform the evaluation:
 ```bash
 # For APPS and CodeContests datasets
-bash /mnt/nas-alinlp/ljy/MoTCoder/MoTCoder/eval/scripts/eval_motcoder_apps_codecontests.sh
+bash eval/scripts/eval_motcoder_apps_codecontests.sh
 
 # For Fix
-bash /mnt/nas-alinlp/ljy/MoTCoder/MoTCoder/eval/scripts/eval_motcoder_fix.sh
+bash eval/scripts/eval_motcoder_fix.sh
 ```
 
 ## Training

eval/scripts/eval_motcoder_apps_codecontests.sh

@@ -4,8 +4,6 @@
 model_list=(
     models/MoTCoder-7B-v1.5
 )
-cd /mnt/nas-alinlp/ljy/MoTCoder/
-
 echo "$model_id/results/chat-template-queries.jsonl"
 
 for model_id in "${model_list[@]}"; do

eval/scripts/eval_motcoder_reflection.sh

@@ -5,9 +5,19 @@ model_list=(
 )
 name=cc-chat-template-queries
 fix_name=cc-chat-template-queries-reflect_n5
-cd /mnt/nas-alinlp/ljy/MoTCoder/
+
 for model_id in "${model_list[@]}"; do
     results_path=$model_id/results/$name-results.jsonl
+
+    python /mnt/nas-alinlp/ljy/MoTCoder/eval/vllm_gen.py \
+        --model_id "$model_id" \
+        --data_path /mnt/nas-alinlp/ljy/MoTCoder/data/prompts/queries_cc.jsonl \
+        --save_path "$model_id/results/$name.jsonl" \
+        --key prompt \
+        --batch_size 1000 \
+        --apply_chat_template \
+        --tensor_parallel_size 8
+
     python /mnt/nas-alinlp/ljy/MoTCoder/eval/eval.py \
         --data_path "$model_id/results/$name.jsonl" \
         --save_path "$model_id/results/$name-metrics.jsonl"  \

imgs/apps.png

@@ -0,0 +1,23 @@
+model_name_or_path: /mnt/nas-alinlp/ljy/models/Qwen2.5-Coder-32B-Instruct
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z3_config.json
+dataset: motcode
+template: qwen
+cutoff_len: 2048
+overwrite_cache: true
+preprocessing_num_workers: 16
+output_dir: /mnt/nas-alinlp/ljy/MoTCoder/output/Qwen2.5-Coder-32B-Instruct-sft-2e-06-mot
+logging_steps: 1
+save_strategy: steps
+plot_loss: true
+save_steps: 200
+per_device_train_batch_size: 4
+gradient_accumulation_steps: 4
+learning_rate: 2.0e-06
+num_train_epochs: 1
+lr_scheduler_type: cosine
+warmup_ratio: 0.03
+bf16: true
+ddp_timeout: 180000000

imgs/codecontests.png

@@ -0,0 +1,23 @@
+model_name_or_path: /mnt/nas-alinlp/ljy/models/Qwen2.5-72B-Instruct
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z3_config.json
+dataset: motcode
+template: qwen
+cutoff_len: 2048
+overwrite_cache: true
+preprocessing_num_workers: 16
+output_dir: /mnt/nas-alinlp/ljy/MoTCoder/output/Qwen2.5-72B-Instruct-sft-2e-06-mot
+logging_steps: 1
+save_strategy: steps
+plot_loss: true
+save_steps: 200
+per_device_train_batch_size: 4
+gradient_accumulation_steps: 4
+learning_rate: 2.0e-06
+num_train_epochs: 1
+lr_scheduler_type: cosine
+warmup_ratio: 0.03
+bf16: true
+ddp_timeout: 180000000

imgs/reflection.png

@@ -0,0 +1,2 @@
+
+llamafactory-cli train train/configs/MoTCoder-32B-Instruct-sft-2e-06-mot.yaml

train/configs/MoTCoder-32B-Instruct-sft-2e-06-mot.yaml

@@ -0,0 +1,2 @@
+
+llamafactory-cli train train/configs/MoTCoder-72B-Instruct-sft-2e-06-mot.yaml