README.md

---
language:
- az
- en
license: ms-pl
task_categories:
- text-retrieval
tags:
- retrieval
- azerbaijani
- information-retrieval
- hard-negatives
- reranker
- ms-marco
- dense-retrieval
- colbert
- bi-encoder
- translated
size_categories:
- 1M<n<10M
configs:
- config_name: corpus
  data_files:
  - split: train
    path: corpus/train-*
- config_name: queries
  data_files:
  - split: train
    path: queries/train-*
- config_name: triplets
  data_files:
  - split: train
    path: triplets/train-*
---

# MS MARCO Azerbaijani — Reranked Retrieval Training Dataset

A large-scale passage retrieval training dataset in Azerbaijani, built by translating a 3.2M subset of the [MS MARCO](https://microsoft.github.io/msmarco/) passage ranking dataset and rescoring all query-passage pairs with a multilingual cross-encoder reranker.

## Overview

| | Count |
|---|---|
| Passages | 8,473,865 |
| Queries | ~800,000 |
| Triplets | ~3,200,000 |
| Negatives per triplet | up to 31 |
| Total pairs scored | 41,746,530 |

## Dataset Configs

The dataset consists of three configs:

### `corpus`

The full translated passage collection.

| Column | Type | Description |
|---|---|---|
| `pid` | int | Passage ID (original MS MARCO pid) |
| `passage` | string | Passage text translated to Azerbaijani |

### `queries`

Translated queries.

| Column | Type | Description |
|---|---|---|
| `qid` | int | Query ID (original MS MARCO qid) |
| `query` | string | Query text translated to Azerbaijani |

### `triplets`

Training triplets with both original MS MARCO scores and reranker scores computed on the Azerbaijani translations.

| Column | Type | Description |
|---|---|---|
| `qid` | int | Query ID (links to `queries`) |
| `pos_pid` | int | Positive passage ID (links to `corpus`) |
| `pos_score_original` | float | Original MS MARCO cross-encoder score (English) |
| `pos_score_reranker` | float | Reranker score on Azerbaijani translation |
| `neg_count` | int | Number of valid negatives for this triplet |
| `neg_{k}_pid` | int | Passage ID of the k-th hard negative |
| `neg_{k}_score_original` | float | Original MS MARCO score of the k-th negative |
| `neg_{k}_score_reranker` | float | Reranker score of the k-th negative (Azerbaijani) |

Negatives are sorted by `score_reranker` descending (hardest first). Columns run from `neg_1_*` to `neg_31_*`.

## Construction Pipeline

1. **Sampling**: 3.2M triplets were sampled from the MS MARCO `examples.json` using reservoir sampling, with 31 negatives selected per query
2. **Translation**: All queries and passages were translated from English to Azerbaijani
3. **Reranking**: Every query-passage pair (positive + all negatives) was scored with [BAAI/bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3) on the Azerbaijani translations (~14 hours, 41.7M pairs scored)
4. **Output**: Triplets with dual scores (original English + Azerbaijani reranker) to enable flexible filtering during training

## Why Reranker Scores?

The original MS MARCO scores were computed on English text. After translation, semantic relationships between queries and passages can shift — some negatives become closer to the positive, and some positives become weaker. The reranker scores on Azerbaijani text reflect what the model will actually see during training.

This also enables **false negative filtering**: negatives with `score_reranker > threshold * pos_score_reranker` are likely correct answers that MS MARCO did not annotate. These can be filtered out during training to avoid noisy supervision signals.

## Usage

```python
from datasets import load_dataset

corpus = load_dataset("LocalDoc/msmarco-az-reranked", "corpus")["train"]
queries = load_dataset("LocalDoc/msmarco-az-reranked", "queries")["train"]
triplets = load_dataset("LocalDoc/msmarco-az-reranked", "triplets")["train"]

# Build lookups
passage_lookup = {row["pid"]: row["passage"] for row in corpus}
query_lookup = {row["qid"]: row["query"] for row in queries}

# Inspect a triplet
t = triplets[0]
print(f"Query: {query_lookup[t['qid']]}")
print(f"Positive [reranker={t['pos_score_reranker']:.4f}]: {passage_lookup[t['pos_pid']][:200]}")

for k in range(1, 4):
    neg_pid = t[f"neg_{k}_pid"]
    neg_score = t[f"neg_{k}_score_reranker"]
    if neg_pid:
        print(f"Neg-{k} [reranker={neg_score:.4f}]: {passage_lookup[neg_pid][:200]}")
```

### Training with False Negative Filtering

```python
# Filter out false negatives where negative score > 95% of positive score
FN_THRESHOLD = 0.95

t = triplets[0]
pos_score = t["pos_score_reranker"]
cutoff = FN_THRESHOLD * pos_score

clean_negs = []
for k in range(1, 32):
    neg_pid = t[f"neg_{k}_pid"]
    neg_score = t[f"neg_{k}_score_reranker"]
    if neg_pid and neg_score < cutoff:
        clean_negs.append((neg_pid, neg_score))

print(f"Original negatives: {t['neg_count']}")
print(f"After FN filtering: {len(clean_negs)}")
```

## Example Output

```
Query: Dişi aslanlar nə qədər doğurur

Positive [original=10.41, reranker=5.64]:
Dişi şir normalda hər 18-26 aydan bir doğur. Təxminən 100-119 günlük
hamiləlik dövründən sonra bir-altı bala doğur. Lakin, balaların sayı
adətən üç və ya dörd olur və hər birinin çəkisi təxminən 3 funt olur.

Neg-1 [original=9.26, reranker=7.41]:  ← false negative (reranker > positive)
Dişi aslanlar adətən hər iki ildən bir bala doğurlar. Dişilər hamilə
və ya əmizdirən deyillərsə, ildə bir neçə dəfə cütləşməyə hazırdırlar.

Neg-2 [original=9.35, reranker=5.41]:
Pride-ın dişi hissəsi bütün yetkinlik həyatlarını birlikdə yaşayır,
lakin erkəklər gəlib-gedir. Dişi aslanın hamiləliyi təxminən dörd ay
davam edir.

Neg-3 [original=3.27, reranker=2.77]:  ← true negative
At: Dişilərin hamiləliyi adətən 11-12 ay çəkir. Dəniz aslanı: Dəniz
şirləri də balalarını 11-12 aylıq hamiləlik dövründən sonra dünyaya
gətirirlər.
```


## Limitations

- Passages and queries are machine-translated; translation artifacts (lexical mismatch, semantic drift) may affect quality
- Reranker scores are from a multilingual model that may underperform on Azerbaijani compared to English
- Original MS MARCO annotations are incomplete — some "negatives" are actually relevant (false negatives)



## Contact

For questions or issues, please contact LocalDoc at [v.resad.89@gmail.com].