from __future__ import annotations import json import logging import os from sentence_transformers import CrossEncoder, SentenceTransformer from mteb import MTEB logging.basicConfig(level=logging.INFO) def test_mteb_rerank(): # Test that reranking works # unfortunately, we need all the query ids to pretend to have this scifact_keys = [ "1", "3", "5", "13", "36", "42", "48", "49", "50", "51", "53", "54", "56", "57", "70", "72", "75", "94", "99", "100", "113", "115", "118", "124", "127", "128", "129", "130", "132", "133", "137", "141", "142", "143", "146", "148", "163", "171", "179", "180", "183", "185", "198", "208", "212", "213", "216", "217", "218", "219", "230", "232", "233", "236", "237", "238", "239", "248", "249", "261", "268", "269", "274", "275", "279", "294", "295", "298", "300", "303", "312", "314", "324", "327", "338", "343", "350", "354", "362", "380", "384", "385", "386", "388", "399", "410", "411", "415", "421", "431", "436", "437", "439", "440", "443", "452", "475", "478", "491", "501", "502", "507", "508", "513", "514", "516", "517", "521", "525", "527", "528", "532", "533", "535", "536", "539", "540", "544", "549", "551", "552", "554", "560", "569", "575", "577", "578", "587", "589", "593", "597", "598", "613", "619", "623", "628", "636", "637", "641", "644", "649", "659", "660", "674", "684", "690", "691", "692", "693", "700", "702", "715", "716", "718", "721", "723", "727", "728", "729", "742", "743", "744", "756", "759", "768", "770", "775", "781", "783", "784", "785", "793", "800", "805", "808", "811", "814", "820", "821", "823", "830", "831", "832", "834", "837", "839", "845", "847", "852", "859", "870", "873", "879", "880", "882", "887", "903", "904", "907", "911", "913", "914", "921", "922", "936", "956", "957", "960", "967", "971", "975", "982", "985", "993", "1012", "1014", "1019", "1020", "1021", "1024", "1029", "1041", "1049", "1062", "1086", "1088", "1089", "1099", "1100", "1104", "1107", "1110", "1121", "1130", "1132", "1137", "1140", "1144", "1146", "1150", "1163", "1175", "1179", "1180", "1185", "1187", "1191", "1194", "1196", "1197", "1199", "1200", "1202", "1204", "1207", "1213", "1216", "1221", "1225", "1226", "1232", "1241", "1245", "1259", "1262", "1266", "1270", "1271", "1272", "1273", "1274", "1278", "1279", "1280", "1281", "1282", "1290", "1292", "1298", "1303", "1316", "1319", "1320", "1332", "1335", "1336", "1337", "1339", "1344", "1352", "1359", "1362", "1363", "1368", "1370", "1379", "1382", "1385", "1389", "1395", ] model = CrossEncoder("cross-encoder/ms-marco-TinyBERT-L-2-v2") eval = MTEB( tasks=[ "SciFact", ] ) # create fake first stage results with open("tmp.json", "w") as f: f.write( json.dumps( { i: { # just two random documents so we can see it works "4983": 0.1, "18670": 0.9, "19238": 0.01, } for i in scifact_keys } ) ) eval.run( model, output_folder="tests/results", overwrite_results=True, eval_splits=["test"], top_k=2, previous_results="tmp.json", save_predictions=True, ) os.remove("tmp.json") # read in the results with open("tests/results/SciFact_default_predictions.json") as f: results = json.load(f) # check that only the top two results are re-orderd assert "19238" not in results["1"] assert "4983" in results["1"] assert "18670" in results["1"] def test_reranker_same_ndcg1(): de = SentenceTransformer("average_word_embeddings_komninos") ce = CrossEncoder("cross-encoder/ms-marco-TinyBERT-L-2-v2") eval = MTEB(tasks=["SciFact"]) eval.run( de, output_folder="tests/results/stage1", overwrite_results=True, save_predictions=True, eval_splits=["test"], ) eval.run( ce, output_folder="tests/results/stage2", overwrite_results=True, previous_results="tests/results/stage1/SciFact_default_predictions.json", save_predictions=False, eval_splits=["test"], top_k=1, # don't allow it to rerank more than 1 so we can check for top_1 being the same ) # read in stage 1 and stage two and check ndcg@1 is the same with open("tests/results/stage1/SciFact.json") as f: stage1 = json.load(f) with open("tests/results/stage2/SciFact.json") as f: stage2 = json.load(f) assert ( stage1["scores"]["test"][0]["ndcg_at_1"] == stage2["scores"]["test"][0]["ndcg_at_1"] )