# Complete FIQA Q4_K_M evaluation

Model: `KaLM-Embedding/KaLM-Reranker-V1-Small-Q4_K_M-GGUF`

## Scope and ranking contract

- BEIR FiQA complete test set: 648 queries.
- Frozen retriever top-100: 64,800 query/passage pairs.
- Rank by margin descending, then retriever rank and passage ID.
- Encoder/query limits: 1,024/512 tokens; encoder chunk size: 4.

## Metrics

| Model | NDCG@10 | MAP@10 | Recall@10 | MRR@10 | NDCG@100 |
| --- | ---: | ---: | ---: | ---: | ---: |
| Retriever | 0.46985 | 0.38419 | 0.55188 | 0.54942 | 0.53769 |
| Transformers/BF16 | 0.55255 | 0.46851 | 0.62310 | 0.63469 | 0.60417 |
| GGUF Q4_K_M | 0.55208 | 0.46747 | 0.62503 | 0.63452 | 0.60340 |

The Q4_K_M gate allows a maximum `0.020` BF16 drop for both NDCG@10 and
MRR@10. Observed drops are `0.00047` and
`0.00017`. Gate status: **accepted**.

Recorded throughput: `26.634`
pairs/s across 81 completed
shards. This number is hardware- and scheduling-specific.

## Provenance

The release manifest records SHA256 values for the source evaluation
manifest, metrics, efficiency, validation and summary.