Negative Sampling Strategy

Design a negative sampling strategy for training/evaluating our retrieval system.

## Retrieval Task
{{retrieval_task}}

## Current Dataset
{{dataset_description}}

## Evaluation Goals
{{evaluation_goals}}

Develop sampling strategies:

**Easy Negatives**
- Random sampling
- Use cases and limitations

**Medium Negatives**
- BM25 top-k that are irrelevant
- Same-topic different answers

**Hard Negatives**
- Semantic near-misses
- Factually incorrect but plausible
- Cross-encoder filtered

**Mining Implementation**
```python
class NegativeMiner:
    def mine_hard_negatives(self, query: str, positive: str, corpus: List[str], k: int) -> List[str]:
        pass
```

Include:
- Ratio recommendations
- Quality filtering
- Curriculum learning approach