--- tags: - mteb model-index: - name: piccolo-base-zh results: - task: type: STS dataset: type: C-MTEB/AFQMC name: MTEB AFQMC config: default split: validation revision: None metrics: - type: cos_sim_pearson value: 49.16558217326158 - type: cos_sim_spearman value: 51.4049475858823 - type: euclidean_pearson value: 49.85853741070363 - type: euclidean_spearman value: 51.501428092542234 - type: manhattan_pearson value: 49.746099634926296 - type: manhattan_spearman value: 51.41081804320127 - task: type: STS dataset: type: C-MTEB/ATEC name: MTEB ATEC config: default split: test revision: None metrics: - type: cos_sim_pearson value: 52.385361699031854 - type: cos_sim_spearman value: 52.59114913702212 - type: euclidean_pearson value: 54.994530439418355 - type: euclidean_spearman value: 52.54102886188004 - type: manhattan_pearson value: 54.9503071669608 - type: manhattan_spearman value: 52.51465652540901 - task: type: Classification dataset: type: mteb/amazon_reviews_multi name: MTEB AmazonReviewsClassification (zh) config: zh split: test revision: 1399c76144fd37290681b995c656ef9b2e06e26d metrics: - type: accuracy value: 40.236 - type: f1 value: 39.43040092463147 - task: type: STS dataset: type: C-MTEB/BQ name: MTEB BQ config: default split: test revision: None metrics: - type: cos_sim_pearson value: 60.98952187211432 - type: cos_sim_spearman value: 62.68189713123115 - type: euclidean_pearson value: 61.089426749761344 - type: euclidean_spearman value: 62.41743375544581 - type: manhattan_pearson value: 61.14747216341409 - type: manhattan_spearman value: 62.488918956547046 - task: type: Clustering dataset: type: C-MTEB/CLSClusteringP2P name: MTEB CLSClusteringP2P config: default split: test revision: None metrics: - type: v_measure value: 38.36392300667918 - task: type: Clustering dataset: type: C-MTEB/CLSClusteringS2S name: MTEB CLSClusteringS2S config: default split: test revision: None metrics: - type: v_measure value: 35.645927581489175 - task: type: Reranking dataset: type: C-MTEB/CMedQAv1-reranking name: MTEB CMedQAv1 config: default split: test revision: None metrics: - type: map value: 85.25085782849087 - type: mrr value: 87.77154761904762 - task: type: Reranking dataset: type: C-MTEB/CMedQAv2-reranking name: MTEB CMedQAv2 config: default split: test revision: None metrics: - type: map value: 86.15357754080844 - type: mrr value: 88.53547619047617 - task: type: Retrieval dataset: type: C-MTEB/CmedqaRetrieval name: MTEB CmedqaRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 23.683 - type: map_at_10 value: 35.522999999999996 - type: map_at_100 value: 37.456 - type: map_at_1000 value: 37.576 - type: map_at_3 value: 31.584 - type: map_at_5 value: 33.684999999999995 - type: mrr_at_1 value: 36.459 - type: mrr_at_10 value: 44.534 - type: mrr_at_100 value: 45.6 - type: mrr_at_1000 value: 45.647 - type: mrr_at_3 value: 42.186 - type: mrr_at_5 value: 43.482 - type: ndcg_at_1 value: 36.459 - type: ndcg_at_10 value: 42.025 - type: ndcg_at_100 value: 49.754 - type: ndcg_at_1000 value: 51.815999999999995 - type: ndcg_at_3 value: 37.056 - type: ndcg_at_5 value: 38.962 - type: precision_at_1 value: 36.459 - type: precision_at_10 value: 9.485000000000001 - type: precision_at_100 value: 1.567 - type: precision_at_1000 value: 0.183 - type: precision_at_3 value: 21.13 - type: precision_at_5 value: 15.209 - type: recall_at_1 value: 23.683 - type: recall_at_10 value: 52.190999999999995 - type: recall_at_100 value: 84.491 - type: recall_at_1000 value: 98.19600000000001 - type: recall_at_3 value: 37.09 - type: recall_at_5 value: 43.262 - task: type: PairClassification dataset: type: C-MTEB/CMNLI name: MTEB Cmnli config: default split: validation revision: None metrics: - type: cos_sim_accuracy value: 74.20324714371618 - type: cos_sim_ap value: 82.32631646194994 - type: cos_sim_f1 value: 76.64052827073876 - type: cos_sim_precision value: 68.58725761772854 - type: cos_sim_recall value: 86.83656768763151 - type: dot_accuracy value: 70.33072760072159 - type: dot_ap value: 77.46972172609794 - type: dot_f1 value: 73.6668924804026 - type: dot_precision value: 62.84676354029062 - type: dot_recall value: 88.98760813654431 - type: euclidean_accuracy value: 74.78051713770296 - type: euclidean_ap value: 82.65778389584023 - type: euclidean_f1 value: 77.1843623157445 - type: euclidean_precision value: 71.05211406096362 - type: euclidean_recall value: 84.47509936871639 - type: manhattan_accuracy value: 74.76849067949489 - type: manhattan_ap value: 82.55694030572194 - type: manhattan_f1 value: 77.1776459569154 - type: manhattan_precision value: 69.5423855963991 - type: manhattan_recall value: 86.69628244096329 - type: max_accuracy value: 74.78051713770296 - type: max_ap value: 82.65778389584023 - type: max_f1 value: 77.1843623157445 - task: type: Retrieval dataset: type: C-MTEB/CovidRetrieval name: MTEB CovidRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 72.99799999999999 - type: map_at_10 value: 81.271 - type: map_at_100 value: 81.53399999999999 - type: map_at_1000 value: 81.535 - type: map_at_3 value: 80.049 - type: map_at_5 value: 80.793 - type: mrr_at_1 value: 73.13 - type: mrr_at_10 value: 81.193 - type: mrr_at_100 value: 81.463 - type: mrr_at_1000 value: 81.464 - type: mrr_at_3 value: 80.067 - type: mrr_at_5 value: 80.741 - type: ndcg_at_1 value: 73.34 - type: ndcg_at_10 value: 84.503 - type: ndcg_at_100 value: 85.643 - type: ndcg_at_1000 value: 85.693 - type: ndcg_at_3 value: 82.135 - type: ndcg_at_5 value: 83.401 - type: precision_at_1 value: 73.34 - type: precision_at_10 value: 9.536 - type: precision_at_100 value: 1.004 - type: precision_at_1000 value: 0.101 - type: precision_at_3 value: 29.54 - type: precision_at_5 value: 18.398 - type: recall_at_1 value: 72.99799999999999 - type: recall_at_10 value: 94.31 - type: recall_at_100 value: 99.368 - type: recall_at_1000 value: 99.789 - type: recall_at_3 value: 87.935 - type: recall_at_5 value: 90.991 - task: type: Retrieval dataset: type: C-MTEB/DuRetrieval name: MTEB DuRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 26.537 - type: map_at_10 value: 81.292 - type: map_at_100 value: 84.031 - type: map_at_1000 value: 84.066 - type: map_at_3 value: 56.571000000000005 - type: map_at_5 value: 71.082 - type: mrr_at_1 value: 91.2 - type: mrr_at_10 value: 93.893 - type: mrr_at_100 value: 93.955 - type: mrr_at_1000 value: 93.95700000000001 - type: mrr_at_3 value: 93.61699999999999 - type: mrr_at_5 value: 93.767 - type: ndcg_at_1 value: 91.2 - type: ndcg_at_10 value: 88.255 - type: ndcg_at_100 value: 90.813 - type: ndcg_at_1000 value: 91.144 - type: ndcg_at_3 value: 87.435 - type: ndcg_at_5 value: 85.961 - type: precision_at_1 value: 91.2 - type: precision_at_10 value: 42.14 - type: precision_at_100 value: 4.817 - type: precision_at_1000 value: 0.48900000000000005 - type: precision_at_3 value: 78.467 - type: precision_at_5 value: 65.75999999999999 - type: recall_at_1 value: 26.537 - type: recall_at_10 value: 89.262 - type: recall_at_100 value: 97.783 - type: recall_at_1000 value: 99.49799999999999 - type: recall_at_3 value: 58.573 - type: recall_at_5 value: 75.154 - task: type: Retrieval dataset: type: C-MTEB/EcomRetrieval name: MTEB EcomRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 48.5 - type: map_at_10 value: 57.898 - type: map_at_100 value: 58.599000000000004 - type: map_at_1000 value: 58.616 - type: map_at_3 value: 55.1 - type: map_at_5 value: 56.80500000000001 - type: mrr_at_1 value: 48.5 - type: mrr_at_10 value: 57.898 - type: mrr_at_100 value: 58.599000000000004 - type: mrr_at_1000 value: 58.616 - type: mrr_at_3 value: 55.1 - type: mrr_at_5 value: 56.80500000000001 - type: ndcg_at_1 value: 48.5 - type: ndcg_at_10 value: 62.876 - type: ndcg_at_100 value: 66.00200000000001 - type: ndcg_at_1000 value: 66.467 - type: ndcg_at_3 value: 57.162 - type: ndcg_at_5 value: 60.263999999999996 - type: precision_at_1 value: 48.5 - type: precision_at_10 value: 7.870000000000001 - type: precision_at_100 value: 0.927 - type: precision_at_1000 value: 0.096 - type: precision_at_3 value: 21.032999999999998 - type: precision_at_5 value: 14.14 - type: recall_at_1 value: 48.5 - type: recall_at_10 value: 78.7 - type: recall_at_100 value: 92.7 - type: recall_at_1000 value: 96.39999999999999 - type: recall_at_3 value: 63.1 - type: recall_at_5 value: 70.7 - task: type: Classification dataset: type: C-MTEB/IFlyTek-classification name: MTEB IFlyTek config: default split: validation revision: None metrics: - type: accuracy value: 44.34782608695652 - type: f1 value: 36.401426200836205 - task: type: Classification dataset: type: C-MTEB/JDReview-classification name: MTEB JDReview config: default split: test revision: None metrics: - type: accuracy value: 84.25891181988743 - type: ap value: 50.54636280166089 - type: f1 value: 78.55080202541332 - task: type: STS dataset: type: C-MTEB/LCQMC name: MTEB LCQMC config: default split: test revision: None metrics: - type: cos_sim_pearson value: 70.02878561337955 - type: cos_sim_spearman value: 75.39509553139982 - type: euclidean_pearson value: 73.92598696939956 - type: euclidean_spearman value: 75.5471147196853 - type: manhattan_pearson value: 73.88049486090739 - type: manhattan_spearman value: 75.51361990583285 - task: type: Retrieval dataset: type: C-MTEB/MMarcoRetrieval name: MTEB MMarcoRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 64.739 - type: map_at_10 value: 74.039 - type: map_at_100 value: 74.38 - type: map_at_1000 value: 74.39099999999999 - type: map_at_3 value: 72.074 - type: map_at_5 value: 73.29299999999999 - type: mrr_at_1 value: 66.92 - type: mrr_at_10 value: 74.636 - type: mrr_at_100 value: 74.94 - type: mrr_at_1000 value: 74.95 - type: mrr_at_3 value: 72.911 - type: mrr_at_5 value: 73.981 - type: ndcg_at_1 value: 66.92 - type: ndcg_at_10 value: 77.924 - type: ndcg_at_100 value: 79.471 - type: ndcg_at_1000 value: 79.73400000000001 - type: ndcg_at_3 value: 74.17200000000001 - type: ndcg_at_5 value: 76.236 - type: precision_at_1 value: 66.92 - type: precision_at_10 value: 9.5 - type: precision_at_100 value: 1.027 - type: precision_at_1000 value: 0.105 - type: precision_at_3 value: 27.989000000000004 - type: precision_at_5 value: 17.874000000000002 - type: recall_at_1 value: 64.739 - type: recall_at_10 value: 89.324 - type: recall_at_100 value: 96.342 - type: recall_at_1000 value: 98.38900000000001 - type: recall_at_3 value: 79.378 - type: recall_at_5 value: 84.28099999999999 - task: type: Classification dataset: type: mteb/amazon_massive_intent name: MTEB MassiveIntentClassification (zh-CN) config: zh-CN split: test revision: 31efe3c427b0bae9c22cbb560b8f15491cc6bed7 metrics: - type: accuracy value: 68.97108271687962 - type: f1 value: 66.8625981386677 - task: type: Classification dataset: type: mteb/amazon_massive_scenario name: MTEB MassiveScenarioClassification (zh-CN) config: zh-CN split: test revision: 7d571f92784cd94a019292a1f45445077d0ef634 metrics: - type: accuracy value: 73.32212508406187 - type: f1 value: 73.33875034670166 - task: type: Retrieval dataset: type: C-MTEB/MedicalRetrieval name: MTEB MedicalRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 49.0 - type: map_at_10 value: 55.022999999999996 - type: map_at_100 value: 55.550999999999995 - type: map_at_1000 value: 55.608000000000004 - type: map_at_3 value: 53.417 - type: map_at_5 value: 54.372 - type: mrr_at_1 value: 49.3 - type: mrr_at_10 value: 55.176 - type: mrr_at_100 value: 55.703 - type: mrr_at_1000 value: 55.76 - type: mrr_at_3 value: 53.567 - type: mrr_at_5 value: 54.522000000000006 - type: ndcg_at_1 value: 49.0 - type: ndcg_at_10 value: 58.089999999999996 - type: ndcg_at_100 value: 60.988 - type: ndcg_at_1000 value: 62.580999999999996 - type: ndcg_at_3 value: 54.803000000000004 - type: ndcg_at_5 value: 56.508 - type: precision_at_1 value: 49.0 - type: precision_at_10 value: 6.78 - type: precision_at_100 value: 0.8210000000000001 - type: precision_at_1000 value: 0.095 - type: precision_at_3 value: 19.6 - type: precision_at_5 value: 12.58 - type: recall_at_1 value: 49.0 - type: recall_at_10 value: 67.80000000000001 - type: recall_at_100 value: 82.1 - type: recall_at_1000 value: 94.8 - type: recall_at_3 value: 58.8 - type: recall_at_5 value: 62.9 - task: type: Reranking dataset: type: C-MTEB/Mmarco-reranking name: MTEB MMarcoReranking config: default split: dev revision: None metrics: - type: map value: 28.87237408060796 - type: mrr value: 27.83015873015873 - task: type: Classification dataset: type: C-MTEB/MultilingualSentiment-classification name: MTEB MultilingualSentiment config: default split: validation revision: None metrics: - type: accuracy value: 70.25 - type: f1 value: 70.29055400149645 - task: type: PairClassification dataset: type: C-MTEB/OCNLI name: MTEB Ocnli config: default split: validation revision: None metrics: - type: cos_sim_accuracy value: 65.56578234975636 - type: cos_sim_ap value: 70.89354058570412 - type: cos_sim_f1 value: 71.21024370095002 - type: cos_sim_precision value: 58.48032564450475 - type: cos_sim_recall value: 91.02428722280888 - type: dot_accuracy value: 64.86193827828912 - type: dot_ap value: 70.17697803463875 - type: dot_f1 value: 70.68676716917922 - type: dot_precision value: 58.57043719639139 - type: dot_recall value: 89.1235480464625 - type: euclidean_accuracy value: 64.86193827828912 - type: euclidean_ap value: 70.26847152773904 - type: euclidean_f1 value: 70.9984152139461 - type: euclidean_precision value: 56.81674064679771 - type: euclidean_recall value: 94.61457233368532 - type: manhattan_accuracy value: 65.40335679480238 - type: manhattan_ap value: 70.22941558736018 - type: manhattan_f1 value: 71.09712937475423 - type: manhattan_precision value: 56.64160401002506 - type: manhattan_recall value: 95.45934530095037 - type: max_accuracy value: 65.56578234975636 - type: max_ap value: 70.89354058570412 - type: max_f1 value: 71.21024370095002 - task: type: Classification dataset: type: C-MTEB/OnlineShopping-classification name: MTEB OnlineShopping config: default split: test revision: None metrics: - type: accuracy value: 89.92999999999999 - type: ap value: 87.16059195012956 - type: f1 value: 89.90917477839415 - task: type: STS dataset: type: C-MTEB/PAWSX name: MTEB PAWSX config: default split: test revision: None metrics: - type: cos_sim_pearson value: 27.74161502387672 - type: cos_sim_spearman value: 31.58353529723325 - type: euclidean_pearson value: 32.43729673844635 - type: euclidean_spearman value: 31.59527486602242 - type: manhattan_pearson value: 32.37467059678786 - type: manhattan_spearman value: 31.44408004951894 - task: type: STS dataset: type: C-MTEB/QBQTC name: MTEB QBQTC config: default split: test revision: None metrics: - type: cos_sim_pearson value: 36.233749845501194 - type: cos_sim_spearman value: 36.47808586229587 - type: euclidean_pearson value: 32.663447466546806 - type: euclidean_spearman value: 34.45830454037139 - type: manhattan_pearson value: 32.80239212096335 - type: manhattan_spearman value: 34.581060433895125 - task: type: STS dataset: type: mteb/sts22-crosslingual-sts name: MTEB STS22 (zh) config: zh split: test revision: None metrics: - type: cos_sim_pearson value: 63.05131937664673 - type: cos_sim_spearman value: 66.51353746725948 - type: euclidean_pearson value: 61.24016998745561 - type: euclidean_spearman value: 66.07115266049276 - type: manhattan_pearson value: 64.55660243659054 - type: manhattan_spearman value: 66.80282149562386 - task: type: STS dataset: type: C-MTEB/STSB name: MTEB STSB config: default split: test revision: None metrics: - type: cos_sim_pearson value: 70.45533692882996 - type: cos_sim_spearman value: 70.6045637565602 - type: euclidean_pearson value: 72.75588977483554 - type: euclidean_spearman value: 73.36630581886473 - type: manhattan_pearson value: 72.72517409326954 - type: manhattan_spearman value: 73.35358940437355 - task: type: Reranking dataset: type: C-MTEB/T2Reranking name: MTEB T2Reranking config: default split: dev revision: None metrics: - type: map value: 66.45779474032288 - type: mrr value: 76.0782192023729 - task: type: Retrieval dataset: type: C-MTEB/T2Retrieval name: MTEB T2Retrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 26.458 - type: map_at_10 value: 74.355 - type: map_at_100 value: 78.158 - type: map_at_1000 value: 78.233 - type: map_at_3 value: 52.2 - type: map_at_5 value: 64.14 - type: mrr_at_1 value: 88.37 - type: mrr_at_10 value: 91.117 - type: mrr_at_100 value: 91.231 - type: mrr_at_1000 value: 91.23599999999999 - type: mrr_at_3 value: 90.645 - type: mrr_at_5 value: 90.948 - type: ndcg_at_1 value: 88.37 - type: ndcg_at_10 value: 82.384 - type: ndcg_at_100 value: 86.431 - type: ndcg_at_1000 value: 87.163 - type: ndcg_at_3 value: 83.993 - type: ndcg_at_5 value: 82.411 - type: precision_at_1 value: 88.37 - type: precision_at_10 value: 41.131 - type: precision_at_100 value: 4.9799999999999995 - type: precision_at_1000 value: 0.515 - type: precision_at_3 value: 73.651 - type: precision_at_5 value: 61.634 - type: recall_at_1 value: 26.458 - type: recall_at_10 value: 81.3 - type: recall_at_100 value: 94.342 - type: recall_at_1000 value: 98.103 - type: recall_at_3 value: 54.020999999999994 - type: recall_at_5 value: 67.781 - task: type: Classification dataset: type: C-MTEB/TNews-classification name: MTEB TNews config: default split: validation revision: None metrics: - type: accuracy value: 46.814 - type: f1 value: 45.580027683507666 - task: type: Clustering dataset: type: C-MTEB/ThuNewsClusteringP2P name: MTEB ThuNewsClusteringP2P config: default split: test revision: None metrics: - type: v_measure value: 61.43613064816144 - task: type: Clustering dataset: type: C-MTEB/ThuNewsClusteringS2S name: MTEB ThuNewsClusteringS2S config: default split: test revision: None metrics: - type: v_measure value: 53.01838461793776 - task: type: Retrieval dataset: type: C-MTEB/VideoRetrieval name: MTEB VideoRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 59.3 - type: map_at_10 value: 69.158 - type: map_at_100 value: 69.60300000000001 - type: map_at_1000 value: 69.611 - type: map_at_3 value: 67.467 - type: map_at_5 value: 68.432 - type: mrr_at_1 value: 59.199999999999996 - type: mrr_at_10 value: 69.108 - type: mrr_at_100 value: 69.553 - type: mrr_at_1000 value: 69.56099999999999 - type: mrr_at_3 value: 67.417 - type: mrr_at_5 value: 68.382 - type: ndcg_at_1 value: 59.3 - type: ndcg_at_10 value: 73.54 - type: ndcg_at_100 value: 75.652 - type: ndcg_at_1000 value: 75.868 - type: ndcg_at_3 value: 70.074 - type: ndcg_at_5 value: 71.808 - type: precision_at_1 value: 59.3 - type: precision_at_10 value: 8.709999999999999 - type: precision_at_100 value: 0.9690000000000001 - type: precision_at_1000 value: 0.099 - type: precision_at_3 value: 25.867 - type: precision_at_5 value: 16.36 - type: recall_at_1 value: 59.3 - type: recall_at_10 value: 87.1 - type: recall_at_100 value: 96.89999999999999 - type: recall_at_1000 value: 98.6 - type: recall_at_3 value: 77.60000000000001 - type: recall_at_5 value: 81.8 - task: type: Classification dataset: type: C-MTEB/waimai-classification name: MTEB Waimai config: default split: test revision: None metrics: - type: accuracy value: 84.69999999999999 - type: ap value: 66.65020528563207 - type: f1 value: 83.00542769081453 --- ## piccolo-base-zh piccolo是一个通用embedding模型(中文), 由来自商汤科技的通用模型组完成训练。piccolo借鉴了E5以及GTE的训练流程，采用了两阶段的训练方式。 在第一阶段中，我们搜集和爬取了4亿的中文文本对(可视为弱监督文本对数据)，并采用二元组的softmax对比学习损失来优化模型。 在第二阶段中，我们搜集整理了2000万人工标注的中文文本对(精标数据)，并采用带有难负样本的三元组的softmax对比学习损失来帮助模型更好地优化。 目前，我们提供了piccolo-base-zh和piccolo-large-zh两个模型。 piccolo is a general text embedding model(chinese), powered by General Model Group from SenseTime Research. Inspired from E5 and GTE, piccolo is trained using a two stage pipeline. On the first stage, we collect and crawl 400 million weakly supervised Chinese text pairs from the Internet, and train the model with the pair(text and text pos) softmax contrastive loss. On the second stage, we collect 20 million human labeled chinese text pairs dataset, and finetune the model with tiplet (text, text_pos, text_neg) contrastive loss. Currently here we offer two different sizes of models, including piccolo-base-zh, piccolo-large-zh. ## Metric 我们将piccolo与其他的开源embedding模型在CMTEB榜单上进行了比较，请参考CMTEB榜单。我们在eval文件夹中提供了复现结果的脚本。 We compared the performance of the piccolo with other embedding models on the C-MTEB benchmark. please refer to the C-MTEB leaderboard. we provide scripts in "eval" folder for results reproducing. | Model Name | Model Size (GB) | Dimension | Sequence Length | Average (35) | Classification (9) | Clustering (4) | Pair Classification (2) | Reranking (4) | Retrieval (8) | STS (8) | |:----:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| | [**piccolo-large-zh**] | 0.65 | 1024 | 512 | **64.11** | 67.03 | 47.04 | 78.38 | 65.98 | 70.93 | 58.02 | | [bge-large-zh]| 1.3 | 1024| 512 | 63.96 | 68.32 | 48.39 | 78.94 | 65.11 | 71.52 | 54.98 | | [**piccolo-base-zh**]| 0.2 | 768 | 512 | **63.66** | 66.98 | 47.12 | 76.61 | 66.68 | 71.2 | 55.9 | | [bge-large-zh-no-instruct]| 1.3 | 1024 | 512 | 63.4 | 68.58 | 50.01 | 76.77 | 64.9 | 70.54 | 53 | | [bge-base-zh]| 0.41 | 768 | 512 | 62.8 | 67.07 | 47.64 | 77.5 | 64.91 | 69.53 | 54.12 | ## Usage 在sentence-transformer package中可以很容易地调用piccolo模型 ```python # for s2s dataset, you can use piccolo as below # 对于短对短数据集，下面是通用的使用方式 from sentence_transformers import SentenceTransformer sentences = ["数据1", "数据2"] model = SentenceTransformer('sensenova/piccolo-base-zh') embeddings_1 = model.encode(sentences, normalize_embeddings=True) embeddings_2 = model.encode(sentences, normalize_embeddings=True) similarity = embeddings_1 @ embeddings_2.T print(similarity) # for s2p dataset, we recommend to add instruction for passage retrieval # 对于短对长数据集，我们推荐添加instruction，来帮助模型更好地进行检索。 from sentence_transformers import SentenceTransformer queries = ['query_1', 'query_2'] passages = ["doc_1", "doc_2"] model = SentenceTransformer('sensenova/piccolo-base-zh') q_embeddings = model.encode(["查询：" + q for q in queries], normalize_embeddings=True) p_embeddings = model.encode(["结果：" + p for p in passages], normalize_embeddings=True) scores = q_embeddings @ p_embeddings.T ``` ## Training Detail ### pretrain pretrain 通常不需要太大的max length, 推荐128。小的max length用以提高batch size，加快训练速度，从而适应大规模数据。 pretrain 损失我们采用二元组contrastive loss，不加入hard negative, 直接采用inbatch negative，在实际训练中，我们使用了32张40G A100进行训练，单卡的batch size为1024。 Pretrain usually does not require a large max length, and 128 is recommended. A small max length is used to increase batch size and speed up training to adapt to large-scale data. We use binary contrastive loss for pretrain loss, without adding hard negative, and directly use inbatch negative. In actual training, we used 32 40G A100 for training, and the batch size of a single card is 1024. ### finetune finetune 通常会将 max length扩增到512。用以适应更大长度的文本输入，finetune时会多sample S2P的数据，以增强模型在retrieval任务上的性能。 finetune 损失采用三元组contrastive loss，加入hard negative，neg num通常设置为2-7，loss计算方式可以参考GTE里的improved contrastive loss。 注意: 我们给query和passage设置了不同的max length，query的max length始终保持在64。 For finetuning, we usually expands the max length to 512. To adapt to larger length text input, finetune will sample more S2P data to enhance the performance of the model on retrieval tasks. The finetune loss uses triple contrastive loss, adding hard negative. Neg num is usually set to 2-7. The loss calculation method can refer to the improved contrastive loss in GTE. Note: We set different max lengths for query and passage, and the max length of query is always kept at 64. ### Others 1. 减小显存的方式: fp16 + gradient checkpointing + ZERO STAGE1 (stage2 不支持双塔结构下的gradient checkpointing) 相关issue见: https://github.com/microsoft/DeepSpeed/issues/988 2. dataset sampler，我们采用了M3E的dataset sampler，用以保证每个batch里的样本均来自于一个dataset，负样本更有价值。 3. instruction。instruction在我们的实验中对retrieval任务有非常大的性能提升，我们在每个训练样本前都加入'查询: '和'结果: '这样的instruction。 1. The way to reduce memory usage: fp16 + gradient checkpointing + ZERO STAGE1 (stage2 does not support gradient checkpointing under the double-tower structure) For related issues, see: https://github.com/microsoft/DeepSpeed/issues/ 988 2. Dataset sampler, we use M3E's dataset sampler to ensure that the samples in each batch come from a dataset, and negative samples are more valuable. 3. instruction. Instruction has greatly improved the performance of the retrieval task in our experiments. We added instructions like 'query: ' and 'result: ' before each training sample. ## License Piccolo 使用 MIT License，免费商用。 Piccolo use MIT License. It can be used for commercial purposes free of charge. ## Acknowledgement piccolo 由来自商汤科技研究院的通用模型组完成训练，[Jinkin](https://huggingface.co/Jinkin) 完成了代码实现和模型训练， [Jinkin](https://huggingface.co/Jinkin), [CCCCxxx](https://huggingface.co/CCCCxxx) 一起完成了数据搜集、整理和评测工作. 项目由 [Gaomengya](https://huggingface.co/gaomengya) 和 [chaorenwu111](https://huggingface.co/chaorenwu111) 主导。 piccolo is powered by Genral Model group from SenseTime Research. [Jinkin](https://huggingface.co/Jinkin) complete code implementation and model training. [Jinkin](https://huggingface.co/Jinkin), [CCCCxxx](https://huggingface.co/CCCCxxx) completed the data collection、processing and model evaluation together. Project is led by [Gaomengya](https://huggingface.co/gaomengya) and [chaorenwu111](https://huggingface.co/chaorenwu111)