import gradio as gr from openai import OpenAI import json import os import uuid import tempfile from tqdm import tqdm import pandas as pd import numpy as np from collections import Counter import time from zipfile import ZipFile # For Azure OpenAI # openai.api_key = os.environ.get("AZURE_OPENAI_KEY") # openai.api_base = os.environ.get("AZURE_OPENAI_ENDPOINT") # openai.api_type = 'azure' # openai.api_version = os.environ.get("AZURE_OPENAI_API_VERSION") # deployment_id = os.environ.get("AZURE_OPENAI_DEP_ID") # gpt_model = deployment_id prompt = """Compare the ground truth and prediction from AI models, to give a correctness score for the prediction. in the ground truth means it is totally right only when all elements in the ground truth are present in the prediction, and means it is totally right when any one element in the ground truth is present in the prediction. The correctness score is 0.0 (totally wrong), 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, or 1.0 (totally right). Just complete the last space of the correctness score. Question | Ground truth | Prediction | Correctness --- | --- | --- | --- What is x in the equation? | -1 -5 | x = 3 | 0.0 What is x in the equation? | -1 -5 | x = -1 | 0.5 What is x in the equation? | -1 -5 | x = -5 | 0.5 What is x in the equation? | -1 -5 | x = -5 or 5 | 0.5 What is x in the equation? | -1 -5 | x = -1 or x = -5 | 1.0 Can you explain this meme? | This meme is poking fun at the fact that the names of the countries Iceland and Greenland are misleading. Despite its name, Iceland is known for its beautiful green landscapes, while Greenland is mostly covered in ice and snow. The meme is saying that the person has trust issues because the names of these countries do not accurately represent their landscapes. | The meme talks about Iceland and Greenland. It's pointing out that despite their names, Iceland is not very icy and Greenland isn't very green. | 0.4 Can you explain this meme? | This meme is poking fun at the fact that the names of the countries Iceland and Greenland are misleading. Despite its name, Iceland is known for its beautiful green landscapes, while Greenland is mostly covered in ice and snow. The meme is saying that the person has trust issues because the names of these countries do not accurately represent their landscapes. | The meme is using humor to point out the misleading nature of Iceland's and Greenland's names. Iceland, despite its name, has lush green landscapes while Greenland is mostly covered in ice and snow. The text 'This is why I have trust issues' is a playful way to suggest that these contradictions can lead to distrust or confusion. The humor in this meme is derived from the unexpected contrast between the names of the countries and their actual physical characteristics. | 1.0 """ import threading, shutil def schedule_cleanup(paths, delay=600): def _clean(): time.sleep(delay) for p in (paths if isinstance(paths, (list, tuple)) else [paths]): try: if os.path.isdir(p): shutil.rmtree(p, ignore_errors=True) elif os.path.isfile(p): os.remove(p) except: pass threading.Thread(target=_clean, daemon=True).start() def grade(file_obj, key, model, api_base, progress=gr.Progress()): if "mmvet" in model: # use our api key for users key = os.environ.get("AZURE_OPENAI_KEY") api_base = os.environ.get("AZURE_OPENAI_ENDPOINT") client = OpenAI( base_url=api_base.strip() if api_base and api_base.strip() else "https://api.openai.com/v1", api_key=key.strip() ) gpt_model = model workdir = tempfile.mkdtemp(prefix="mmvet_grade_") uid = uuid.uuid4().hex # load metadata # Download mm-vet.zip and `unzip mm-vet.zip` and change the path below mmvet_path = "mm-vet" use_sub_set = False decimal_places = 1 # number of decimal places to round to if use_sub_set: bard_set_file = os.path.join(mmvet_path, "bard_set.json") with open(bard_set_file, 'r') as f: sub_set = json.load(f) sub_set_name = 'bardset' sub_set_name = sub_set_name + '_' else: sub_set = None sub_set_name = '' mmvet_metadata = os.path.join(mmvet_path, "mm-vet.json") with open(mmvet_metadata, 'r') as f: data = json.load(f) counter = Counter() cap_set_list = [] cap_set_counter = [] len_data = 0 for id, value in data.items(): if sub_set is not None and id not in sub_set: continue question = value["question"] answer = value["answer"] cap = value["capability"] cap = set(cap) counter.update(cap) if cap not in cap_set_list: cap_set_list.append(cap) cap_set_counter.append(1) else: cap_set_counter[cap_set_list.index(cap)] += 1 len_data += 1 sorted_list = counter.most_common() columns = [k for k, v in sorted_list] columns.append("total") columns.append("std") columns.append('runs') df = pd.DataFrame(columns=columns) cap_set_sorted_indices = np.argsort(-np.array(cap_set_counter)) new_cap_set_list = [] new_cap_set_counter = [] for index in cap_set_sorted_indices: new_cap_set_list.append(cap_set_list[index]) new_cap_set_counter.append(cap_set_counter[index]) cap_set_list = new_cap_set_list cap_set_counter = new_cap_set_counter cap_set_names = ["_".join(list(cap_set)) for cap_set in cap_set_list] columns2 = cap_set_names columns2.append("total") columns2.append("std") columns2.append('runs') df2 = pd.DataFrame(columns=columns2) ###### change your model name ###### model_name = os.path.basename(file_obj.name)[:-5] # result_path = "results" num_run = 1 # we set 5 in the paper # model_results_file = os.path.join(result_path, f"{model}.json") model_results_file = file_obj.name grade_file = os.path.join(workdir, f'{model_name}_{gpt_model.replace("-mmvet", "")}-grade-{num_run}runs_{uid}.json') cap_score_file = os.path.join(workdir, f'{model_name}_{sub_set_name}{gpt_model.replace("-mmvet", "")}-cap-score-{num_run}runs_{uid}.csv') cap_int_score_file = os.path.join(workdir, f'{model_name}_{sub_set_name}{gpt_model.replace("-mmvet", "")}-cap-int-score-{num_run}runs_{uid}.csv') zip_file = os.path.join(workdir, f"results_{uid}.zip") with open(model_results_file) as f: results = json.load(f) if os.path.exists(grade_file): with open(grade_file, 'r') as f: grade_results = json.load(f) else: grade_results = {} def need_more_runs(): need_more_runs = False if len(grade_results) > 0: for k, v in grade_results.items(): if len(v['score']) < num_run: need_more_runs = True break return need_more_runs or len(grade_results) < len_data while need_more_runs(): for j in range(num_run): print(f'eval run {j}') for id, line in progress.tqdm(data.items(), desc="Grading"): if sub_set is not None and id not in sub_set: continue if id in grade_results and len(grade_results[id]['score']) >= (j + 1): continue model_pred = results[id] question = prompt + '\n' + ' | '.join([line['question'], line['answer'].replace("", " ").replace("", " "), model_pred, ""]) messages = [ {"role": "user", "content": question}, ] if id not in grade_results: sample_grade = {'model': [], 'content': [], 'score': []} else: sample_grade = grade_results[id] grade_sample_run_complete = False temperature = 0.0 num_sleep = 0 while not grade_sample_run_complete: try: response = client.chat.completions.create( model=gpt_model, # engine=gpt_model, # For Azure OpenAI max_tokens=3, temperature=temperature, messages=messages) content = response.choices[0].message.content flag = True try_time = 1 while flag: try: content = content.split(' ')[0].strip() score = float(content) if score > 1.0 or score < 0.0: assert False flag = False except: question = prompt + '\n' + ' | '.join([line['question'], line['answer'].replace("", " ").replace("", " "), model_pred, ""]) + "\nPredict the correctness of the answer (digit): " messages = [ {"role": "user", "content": question}, ] response = client.chat.completions.create( model=gpt_model, # engine=gpt_model, # For Azure OpenAI max_tokens=3, temperature=temperature, messages=messages) content = response.choices[0].message.content try_time += 1 temperature += 0.5 print(f"{id} try {try_time} times") print(content) if try_time > 5: score = 0.0 flag = False grade_sample_run_complete = True except Exception as e: print(e) # gpt4 may have token rate limit num_sleep += 1 if num_sleep > 12: score = 0.0 grade_sample_run_complete = True num_sleep = 0 continue print("sleep 5s") time.sleep(5) resp_model = (getattr(response, "model", None) or gpt_model) content_str = str(content) if len(sample_grade['model']) >= j + 1: sample_grade['model'][j] = resp_model sample_grade['content'][j] = content_str sample_grade['score'][j] = score else: sample_grade['model'].append(resp_model) sample_grade['content'].append(content_str) sample_grade['score'].append(score) grade_results[id] = sample_grade with open(grade_file, 'w') as f: json.dump(grade_results, f, indent=4) assert not need_more_runs() cap_socres = {k: [0.0]*num_run for k in columns[:-2]} counter['total'] = len_data cap_socres2 = {k: [0.0]*num_run for k in columns2[:-2]} counter2 = {columns2[i]:cap_set_counter[i] for i in range(len(cap_set_counter))} counter2['total'] = len_data for k, v in grade_results.items(): if sub_set is not None and k not in sub_set: continue for i in range(num_run): score = v['score'][i] caps = set(data[k]['capability']) for c in caps: cap_socres[c][i] += score cap_socres['total'][i] += score index = cap_set_list.index(caps) cap_socres2[cap_set_names[index]][i] += score cap_socres2['total'][i] += score for k, v in cap_socres.items(): cap_socres[k] = np.array(v) / counter[k] *100 std = round(cap_socres['total'].std(), decimal_places) total_copy = cap_socres['total'].copy() runs = str(list(np.round(total_copy, decimal_places))) for k, v in cap_socres.items(): cap_socres[k] = round(v.mean(), decimal_places) cap_socres['std'] = std cap_socres['runs'] = runs df.loc[gpt_model.replace("-mmvet", "")] = cap_socres for k, v in cap_socres2.items(): cap_socres2[k] = round(np.mean(np.array(v) / counter2[k] *100), decimal_places) cap_socres2['std'] = std cap_socres2['runs'] = runs df2.loc[gpt_model.replace("-mmvet", "")] = cap_socres2 df.to_csv(cap_score_file) df2.to_csv(cap_int_score_file) files = [cap_score_file, cap_int_score_file, grade_file] with ZipFile(zip_file, "w") as zipObj: for fpath in files: arcname = os.path.basename(fpath) zipObj.write(fpath, arcname) for fpath in files: os.remove(fpath) schedule_cleanup([zip_file, workdir], delay=3600) return zip_file # demo = gr.Interface( # fn=grade, # inputs=gr.File(file_types=[".json"]), # outputs="file") # --- Validate key and model before running grading --- def validate_key_and_model(key: str, model: str, api_base: str = None): try: client = OpenAI( base_url=api_base.strip() if api_base and api_base.strip() else "https://api.openai.com/v1", api_key=key.strip() ) client.models.retrieve(model) return True, "OK" except Exception as e: return False, str(e) # --- Wrapper for the grading function --- def run_grade(file_obj, key, model, api_base, progress=gr.Progress(track_tqdm=True)): if model == "gpt-4.1": model = "gpt-4.1-mmvet" # in our Azure OpenAI base, the model name is gpt-4.1-mmvet if "mmvet" not in model: ok, msg = validate_key_and_model(key, model, api_base) if not ok: raise gr.Error(msg) return grade(file_obj, key, model, api_base, progress=progress) markdown = """


# [MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities](https://arxiv.org/abs/2308.02490) This demo uses LLM-based (GPT-4) evaluator to grade open-ended outputs from your models. Plese upload your json file of your model results containing `{v1_0: ..., v1_1: ..., }`like [this json file](https://raw.githubusercontent.com/yuweihao/MM-Vet/main/results/llava_llama2_13b_chat.json). The grading may last 5 minutes. Sine we only support 1 queue, the grading time may be longer when you need to wait for other users' grading to finish. The grading results will be downloaded as a zip file. """ with gr.Blocks() as demo: gr.Markdown(markdown) # Model selection model = gr.Dropdown( choices=["gpt-4.1", "gpt-4-0613", "gpt-4-turbo"], value="gpt-4.1", label="Select model (gpt-4.1 is free with our api key)" ) # User OpenAI fields (only for non-Azure models) with gr.Row(): user_key = gr.Textbox( label="Your OpenAI API Key (required for gpt-4-0613 (default in the paper) / gpt-4-turbo). The evaluation may cost several dollars, please notice your OpenAI API Key balance. 1M input tokens: gpt-4-turbo $10.00, gpt-4-0613 $30.00", type="password", visible=False ) user_api_base = gr.Textbox( label="Your OpenAI Base URL (optional, leave empty for official)", value="", visible=False ) # File I/O with gr.Row(): inp = gr.File(file_types=[".json"], label="Upload your model result JSON") out = gr.File(file_types=[".zip"], label="Download grading results") btn = gr.Button("Start grading", variant="primary") # Toggle fields based on selection def _toggle_fields(selected): if selected == "gpt-4.1": return gr.update(visible=False), gr.update(visible=False) else: return gr.update(visible=True), gr.update(visible=True) model.change(_toggle_fields, inputs=[model], outputs=[user_key, user_api_base]) # Click handler btn.click( fn=run_grade, inputs=[inp, user_key, model, user_api_base], outputs=out ) if __name__ == "__main__": demo.queue(max_size=8).launch()