{"cells":[{"cell_type":"code","execution_count":1,"metadata":{"executionInfo":{"elapsed":476,"status":"ok","timestamp":1720679526275,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"uWKRSV6eZsCn"},"outputs":[],"source":["%load_ext autoreload\n","%autoreload 2"]},{"cell_type":"code","execution_count":3,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["Python 3.11.9\n","Name: transformers\n","Version: 4.41.2\n","Summary: State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow\n","Home-page: https://github.com/huggingface/transformers\n","Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)\n","Author-email: transformers@huggingface.co\n","License: Apache 2.0 License\n","Location: /Users/inflaton/anaconda3/envs/logical-reasoning/lib/python3.11/site-packages\n","Requires: filelock, huggingface-hub, numpy, packaging, pyyaml, regex, requests, safetensors, tokenizers, tqdm\n","Required-by: peft\n","---\n","Name: torch\n","Version: 2.5.0.dev20240720\n","Summary: Tensors and Dynamic neural networks in Python with strong GPU acceleration\n","Home-page: https://pytorch.org/\n","Author: PyTorch Team\n","Author-email: packages@pytorch.org\n","License: BSD-3\n","Location: /Users/inflaton/anaconda3/envs/logical-reasoning/lib/python3.11/site-packages\n","Requires: filelock, fsspec, jinja2, networkx, sympy, typing-extensions\n","Required-by: accelerate, peft, torchaudio, torchvision\n","---\n","Name: torchvision\n","Version: 0.20.0.dev20240721\n","Summary: image and video datasets and models for torch deep learning\n","Home-page: https://github.com/pytorch/vision\n","Author: PyTorch Core Team\n","Author-email: soumith@pytorch.org\n","License: BSD\n","Location: /Users/inflaton/anaconda3/envs/logical-reasoning/lib/python3.11/site-packages\n","Requires: numpy, pillow, torch\n","Required-by: \n","---\n","Name: torchaudio\n","Version: 2.4.0.dev20240721\n","Summary: An audio package for PyTorch\n","Home-page: https://github.com/pytorch/audio\n","Author: Soumith Chintala, David Pollack, Sean Naren, Peter Goldsborough, Moto Hira, Caroline Chen, Jeff Hwang, Zhaoheng Ni, Xiaohui Zhang\n","Author-email: soumith@pytorch.org\n","License: \n","Location: /Users/inflaton/anaconda3/envs/logical-reasoning/lib/python3.11/site-packages\n","Requires: torch\n","Required-by: \n","CPU times: user 8.21 ms, sys: 9.66 ms, total: 17.9 ms\n","Wall time: 2.72 s\n"]}],"source":["%%time\n","!python --version\n","!pip show transformers torch torchvision torchaudio"]},{"cell_type":"code","execution_count":4,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"eb33b19f-1206-41ee-84e2-e6258a12eef7","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":2534,"status":"ok","timestamp":1720679529344,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"xwFh14uiZBrI","outputId":"d767799c-34c2-46a5-f052-378146a55321"},"outputs":[],"source":["from pathlib import Path\n","\n","try:\n"," from google.colab import drive\n","\n"," drive.mount(\"/content/drive\")\n"," workding_dir = \"/content/drive/MyDrive/logical-reasoning/\"\n","except ModuleNotFoundError:\n"," workding_dir = str(Path.cwd().parent)"]},{"cell_type":"code","execution_count":5,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"6d394937-6c99-4a7c-9d32-7600a280032f","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":5,"status":"ok","timestamp":1720679529345,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"G5pNu3zgZBrL","outputId":"160a554f-fb08-4aa0-bc00-0422fb7c1fac"},"outputs":[{"name":"stdout","output_type":"stream","text":["workding dir: /Users/inflaton/code/engd/projects/logical-reasoning\n"]}],"source":["import os\n","import sys\n","from pathlib import Path\n","\n","os.chdir(workding_dir)\n","sys.path.append(workding_dir)\n","print(\"workding dir:\", workding_dir)"]},{"cell_type":"code","execution_count":6,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"9f67ec60-2f24-411c-84eb-0dd664b44775","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":3,"status":"ok","timestamp":1720679529345,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"hPCC-6m7ZBrM","outputId":"c7aa2c96-5e99-440a-c148-201d79465ff9"},"outputs":[{"name":"stdout","output_type":"stream","text":["loading env vars from: /Users/inflaton/code/engd/projects/logical-reasoning/.env\n"]},{"data":{"text/plain":["True"]},"execution_count":6,"metadata":{},"output_type":"execute_result"}],"source":["from dotenv import find_dotenv, load_dotenv\n","\n","found_dotenv = find_dotenv(\".env\")\n","\n","if len(found_dotenv) == 0:\n"," found_dotenv = find_dotenv(\".env.example\")\n","print(f\"loading env vars from: {found_dotenv}\")\n","load_dotenv(found_dotenv, override=True)"]},{"cell_type":"code","execution_count":7,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"f1597656-8042-4878-9d3b-9ebfb8dd86dc","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":3,"status":"ok","timestamp":1720679529345,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"1M3IraVtZBrM","outputId":"29ab35f6-2970-4ade-d85d-3174acf8cda0"},"outputs":[{"name":"stdout","output_type":"stream","text":["internlm/internlm2_5-7b-chat-1m llama-factory/saves/internlm2_5_7b/lora/sft_bf16_p2_full_r2/checkpoint-175 False datasets/mgtv results/mgtv-results_m3.csv\n"]}],"source":["import os\n","\n","model_name = os.getenv(\"MODEL_NAME\")\n","adapter_name_or_path = os.getenv(\"ADAPTER_NAME_OR_PATH\")\n","load_in_4bit = os.getenv(\"LOAD_IN_4BIT\") == \"true\"\n","data_path = os.getenv(\"LOGICAL_REASONING_DATA_PATH\")\n","results_path = os.getenv(\"LOGICAL_REASONING_RESULTS_PATH\")\n","use_english_datasets = os.getenv(\"USE_ENGLISH_DATASETS\") == \"true\"\n","\n","print(model_name, adapter_name_or_path, load_in_4bit, data_path, results_path)"]},{"cell_type":"code","execution_count":8,"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":1685,"status":"ok","timestamp":1720679531591,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"ZuS_FsLyZBrN","outputId":"2cba0105-c505-4395-afbd-2f2fee6581d0"},"outputs":[{"name":"stdout","output_type":"stream","text":["loading /Users/inflaton/code/engd/projects/logical-reasoning/llm_toolkit/logical_reasoning_utils.py\n","MPS is available\n"]}],"source":["from llm_toolkit.llm_utils import *\n","from llm_toolkit.logical_reasoning_utils import *\n","\n","device = check_gpu()"]},{"cell_type":"code","execution_count":9,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["loading model: internlm/internlm2_5-7b-chat-1m with adapter: llama-factory/saves/internlm2_5_7b/lora/sft_bf16_p2_full_r2/checkpoint-175\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"a505972d0ae14d4aab281798aca28c44","version_major":2,"version_minor":0},"text/plain":["Loading checkpoint shards: 0%| | 0/8 [00:00<|im_start|>system\n","You are an expert in logical reasoning.<|im_end|>\n","<|im_start|>user\n","你是一个情景猜谜游戏的主持人。游戏规则如下：\n","\n","1. 参与者会得到一个谜面，谜面会描述一个简单又难以理解的事件。\n","2. 主持人知道谜底，谜底是谜面的答案。\n","3. 参与者可以询问任何封闭式问题来找寻事件的真相。\n","4. 对于每个问题，主持人将根据实际情况回答以下五个选项之一：是、不是、不重要、回答正确、问法错误。各回答的判断标准如下：\n"," - 若谜面和谜底能找到问题的答案，回答：是或者不是\n"," - 若谜面和谜底不能直接或者间接推断出问题的答案，回答：不重要\n"," - 若参与者提问不是一个封闭式问题或者问题难以理解，回答：问法错误\n"," - 若参与者提问基本还原了谜底真相，回答：回答正确\n","5. 回答中不能添加任何其它信息，也不能省略选项中的任何一个字。例如，不可以把“不是”省略成“不”。\n","\n","请严格按照这些规则回答参与者提出的问题。\n","\n","**谜面:** 在一个安静的夜晚，小村庄的湖边突然传来了阵阵哭泣声。第二天早晨，村长甄锐发现湖边的石头上放着一顶破旧的帽子，但没有人知道这顶帽子是从哪里来的，哭泣声又是为何。请还原故事真相。\n","\n","**谜底:** 原来，这顶破旧的帽子属于一个小男孩，他小时候与爷爷在湖边生活。爷爷教他钓鱼、游泳，还告诉他湖中的海龟是他们的朋友。后来，小男孩随父母去了城市生活，但每年夏天都会回到村子探望爷爷。然而，去年夏天，爷爷因病去世，小男孩伤心欲绝。今年夏天，他回到村子，来到湖边，想起和爷爷的美好回忆，忍不住哭泣。他将爷爷的帽子放在湖边的石头上，希望能让爷爷的在天之灵得到安慰。那晚的哭泣声正是小男孩在祭莫他亲爱的爷爷。\n","\n","**参与者提出的问题:** 哭泣和村庄有关系吗\n","<|im_end|>\n","<|im_start|>assistant\n","是\n","--------------------------------------------------\n","prompt: <|im_start|>system\n","You are an expert in logical reasoning.<|im_end|>\n","<|im_start|>user\n","你是一个情景猜谜游戏的主持人。游戏规则如下：\n","\n","1. 参与者会得到一个谜面，谜面会描述一个简单又难以理解的事件。\n","2. 主持人知道谜底，谜底是谜面的答案。\n","3. 参与者可以询问任何封闭式问题来找寻事件的真相。\n","4. 对于每个问题，主持人将根据实际情况回答以下五个选项之一：是、不是、不重要、回答正确、问法错误。各回答的判断标准如下：\n"," - 若谜面和谜底能找到问题的答案，回答：是或者不是\n"," - 若谜面和谜底不能直接或者间接推断出问题的答案，回答：不重要\n"," - 若参与者提问不是一个封闭式问题或者问题难以理解，回答：问法错误\n"," - 若参与者提问基本还原了谜底真相，回答：回答正确\n","5. 回答中不能添加任何其它信息，也不能省略选项中的任何一个字。例如，不可以把“不是”省略成“不”。\n","\n","请严格按照这些规则回答参与者提出的问题。\n","\n","**谜面:** 在一个安静的夜晚，小村庄的湖边突然传来了阵阵哭泣声。第二天早晨，村长甄锐发现湖边的石头上放着一顶破旧的帽子，但没有人知道这顶帽子是从哪里来的，哭泣声又是为何。请还原故事真相。\n","\n","**谜底:** 原来，这顶破旧的帽子属于一个小男孩，他小时候与爷爷在湖边生活。爷爷教他钓鱼、游泳，还告诉他湖中的海龟是他们的朋友。后来，小男孩随父母去了城市生活，但每年夏天都会回到村子探望爷爷。然而，去年夏天，爷爷因病去世，小男孩伤心欲绝。今年夏天，他回到村子，来到湖边，想起和爷爷的美好回忆，忍不住哭泣。他将爷爷的帽子放在湖边的石头上，希望能让爷爷的在天之灵得到安慰。那晚的哭泣声正是小男孩在祭莫他亲爱的爷爷。\n","\n","**参与者提出的问题:** 哭泣和村庄有关系吗\n","<|im_end|>\n","<|im_start|>assistant\n","\n"]}],"source":["print_row_details(datasets[\"test\"].to_pandas(), [1000])"]},{"cell_type":"code","execution_count":12,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["是\n","CPU times: user 619 ms, sys: 392 ms, total: 1.01 s\n","Wall time: 1.9 s\n"]}],"source":["%%time\n","\n","prompt1 = datasets[\"test\"][\"prompt\"][1000]\n","\n","gen_kwargs = {\"max_length\": 4096, \"do_sample\": True, \"top_k\": 1}\n","with torch.no_grad():\n"," inputs = tokenizer(\n"," [prompt1],\n"," return_tensors=\"pt\",\n"," ).to(device)\n"," outputs = model.generate(**inputs, **gen_kwargs)\n"," outputs = outputs[:, inputs['input_ids'].shape[1]:]\n"," print(tokenizer.decode(outputs[0], skip_special_tokens=True))"]},{"cell_type":"code","execution_count":13,"metadata":{},"outputs":[],"source":["def evaluate_model(model, tokenizer, dataset, batch_size=8):\n"," save_model_name = f\"{model_name}_{adapter_name_or_path}_m3_{model.dtype}\"\n"," print(f\"Evaluating model: {save_model_name} on {device}\")\n"," predictions = eval_model(\n"," model, tokenizer, dataset, device=device, batch_size=batch_size\n"," )\n","\n"," save_results(\n"," save_model_name,\n"," results_path,\n"," dataset,\n"," predictions,\n"," debug=False,\n"," )\n","\n"," metrics = calc_metrics(dataset[\"label\"], predictions, debug=False)\n"," print(metrics)"]},{"cell_type":"code","execution_count":14,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["Evaluating model: internlm/internlm2_5-7b-chat-1m_llama-factory/saves/internlm2_5_7b/lora/sft_bf16_p2_full_r2/checkpoint-175_m3_torch.bfloat16 on mps\n"]},{"name":"stderr","output_type":"stream","text":[" 0%| | 1/375 [00:15<1:35:45, 15.36s/it]"]},{"name":"stdout","output_type":"stream","text":["Batch output: ['不是', '是', '是', '是', '不是', '是', '是', '不是']\n"]},{"name":"stderr","output_type":"stream","text":["100%|██████████| 375/375 [2:42:25<00:00, 25.99s/it] \n"]},{"name":"stdout","output_type":"stream","text":["{'accuracy': 0.807}\n","CPU times: user 5min 2s, sys: 5min 33s, total: 10min 35s\n","Wall time: 2h 42min 25s\n"]}],"source":["%%time\n","\n","evaluate_model(model, tokenizer, datasets[\"test\"])"]},{"cell_type":"code","execution_count":15,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["loading model: internlm/internlm2_5-7b-chat-1m with adapter: llama-factory/saves/internlm2_5_7b/lora/sft_bf16_p2_full_r2/checkpoint-175\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"e3301a4dc82449938e9bedb9d8ec5754","version_major":2,"version_minor":0},"text/plain":["Loading checkpoint shards: 0%| | 0/8 [00:00