README.md

---
license: mit
library_name: sentence-transformers
pipeline_tag: text-ranking
tags:
  - cross-encoder
  - reranker
  - climate-fact-checking
  - comp90042
---

# New-Methodology finetuned rerankers

Cross-encoder / reranker checkpoints produced by `scripts/02_train_reranker.py` and related experiments.

## Subfolders

| HF path | Local source |
|---------|--------------|
| `finetuned_reranker/` | `outputs/retrieval/models/finetuned_reranker` |
| `ce_minilm_hybrid_hard_epochs2/` | `outputs/retrieval/models/ce_cross-encoder__ms-marco-MiniLM-L6-v2__hybrid_hard__epochs2` |
| `finetuned_bge_reranker/` | `outputs/retrieval/models/finetuned_bge_reranker` |
| `finetuned_minilm_hardneg/` | `outputs/retrieval/models/finetuned_minilm_hardneg` |
| `finetuned_reranker_s60/` | `outputs/retrieval/models/finetuned_reranker_s60` |
| `pairwise_reranker/` | `outputs/reranker/pairwise/pairwise_reranker` |

## Load example (sentence-transformers)

```python
from sentence_transformers import CrossEncoder

model = CrossEncoder("jasperyeoh2/new-methodology-rerankers", trust_remote_code=True)
# For a specific checkpoint, clone and point model_name_or_path to the subfolder.
```

## Recommended for MMR pipeline

Production MMR (λ=0.5) uses **zero-shot** `cross-encoder/ms-marco-MiniLM-L6-v2`, not these finetuned weights.
Finetuned checkpoints are kept for reproducibility and ablation.