# Dense Retrieval Baseline Summary

All runs use `K=4`, greedy decoding, and `support_input + support_output` as the retrieval passage text.

| Task | Method | Model | Year | ROUGE-L | METEOR | SFD_-len | Avg Len |
|---|---|---|---|---:|---:|---:|---:|
| topic_user | dense_minilm_top1 | sentence-transformers/all-MiniLM-L6-v2 | MiniLM 2020; ST checkpoint circa 2021 | 0.1166 | 0.1866 | 0.7993 | 227.2 |
| topic_user | dense_mpnet_top1 | sentence-transformers/all-mpnet-base-v2 | MPNet 2020; ST checkpoint circa 2021 | 0.1182 | 0.1888 | 0.8180 | 232.7 |
| topic_user | dense_e5_top1 | intfloat/e5-base-v2 | E5 2022 | 0.1191 | 0.1875 | 0.8804 | 236.1 |
| topic_user | dense_bge_top1 | BAAI/bge-base-en-v1.5 | BGE v1.5 2023 | 0.1191 | 0.1857 | 0.7932 | 225.0 |
| review_user | dense_minilm_top1 | sentence-transformers/all-MiniLM-L6-v2 | MiniLM 2020; ST checkpoint circa 2021 | 0.1363 | 0.1901 | 0.6700 | 213.2 |
| review_user | dense_mpnet_top1 | sentence-transformers/all-mpnet-base-v2 | MPNet 2020; ST checkpoint circa 2021 | 0.1374 | 0.1903 | 0.7027 | 217.5 |
| review_user | dense_e5_top1 | intfloat/e5-base-v2 | E5 2022 | 0.1378 | 0.1914 | 0.6782 | 219.2 |
| review_user | dense_bge_top1 | BAAI/bge-base-en-v1.5 | BGE v1.5 2023 | 0.1398 | 0.1923 | 0.6610 | 214.8 |

Best dense retriever:

- `topic_user`: `dense_bge_top1` by ROUGE-L, essentially tied with `dense_e5_top1`.
- `review_user`: `dense_bge_top1`.