nq-mxbai-crossencoder-v1

nq-mxbai-crossencoder-v1

Cross-encoder reranker for the antiquarian reply retrieval task (Notes and Queries 1849-1928, 1 query vs 20 candidate replies, MRR). Ensemble member alongside shikazu05/nq-deberta-crossencoder-v1 (per-case average of blended scores).

Public artifact of a documented fine-tune. No login, token, or payment needed to download or use; inference runs locally via transformers (no external inference API).

Training provenance (reproducible)

• Base: mixedbread-ai/mxbai-rerank-base-v1 (DeBERTa-v3-base 184M, Apache 2.0; itself a public MS-MARCO-style reranker).
• Data (text only): train.csv + train_labels.csv (1,472 cases x 20 candidates = 29,440 query/reply pairs) and replies.csv. Preprocessing: long-s ſ→s, whitespace collapse only.
• Objective: listwise CrossEntropy over each query's 20 candidates.
• Hyperparams: 4 epochs, best epoch 1 (holdout early stop), lr 2e-5 cosine, AdamW wd 0.01, fp16, max_length 320, cases_per_step 1 with gradient accumulation to 4, seed 42. Hardware: 1xA10G (24GB).
• Code: modal_app.py::train_listwise, stage "train-mxbai".

Metrics

• Holdout (147 cases, seed-42 split): frozen 0.4644, ep1 0.6095 (best), ep2-4 ~0.594-0.598. Sibling DeBERTa-raw run: 0.6156.
• Ensemble (DeBERTa-raw + mxbai, TF-IDF blend 0.10, score average): see solution.py header.

Reproduce

modal run modal_app.py --stage train-mxbai --push-hub 0