Cross Encoder Training

Cross-Encoder Training

When to Fine-Tune

Off-the-shelf rerankers (Cohere Rerank v3.5, BGE-reranker-v2-m3, Voyage rerank-2) are strong. Fine-tune when:

• Labeled gold set shows NDCG@10 plateauing around the vendor baseline.
• Your domain vocabulary drifts from web text (legal contracts, clinical notes, code).
• You must self-host a small, fast reranker (100-300 MB) for edge deployment.
• You need to distill a proprietary vendor model into something you own.

Skip fine-tuning if you have fewer than 2k labeled triples. Spend effort on better retrieval instead.

Training Data Shapes

Pointwise

(query, doc, label) where label in {0, 1} or a graded relevance score.

Pairwise

(query, positive_doc, negative_doc) — the model learns to score positive above negative.

Listwise (groups)

(query, [d1, d2, ..., dn], [relevance1, ..., relevance_n]).

MS MARCO uses pairwise triples; most open cross-encoders start from MS MARCO plus domain data.

Minimal CrossEncoder Training (pointwise)

# pip install sentence-transformers==3.* torch
from sentence_transformers import CrossEncoder, InputExample
from sentence_transformers.cross_encoder.losses import BinaryCrossEntropyLoss
from torch.utils.data import DataLoader

train_examples = [
    InputExample(texts=[q, pos], label=1.0) for q, pos in positives
] + [
    InputExample(texts=[q, neg], label=0.0) for q, neg in negatives
]

model = CrossEncoder(
    "cross-encoder/ms-marco-MiniLM-L-6-v2",
    num_labels=1,
    max_length=512,
)

train_loader = DataLoader(train_examples, shuffle=True, batch_size=32)
model.fit(
    train_dataloader=train_loader,
    epochs=2,
    warmup_steps=500,
    output_path="./models/cross-encoder-domain",
    optimizer_params={"lr": 2e-5},
    use_amp=True,
)

MiniLM-L-6 is a 22 M-param backbone — fast at serving (~1-5 ms per pair on T4). For higher quality, start from cross-encoder/ms-marco-MiniLM-L-12-v2 or BAAI/bge-reranker-v2-m3.

Pairwise MarginRankingLoss

from sentence_transformers.cross_encoder.losses import MarginMSELoss

# triples: (query, positive, negative) with teacher scores
triples = [
    InputExample(texts=[q, p, n], label=float(teacher_pos - teacher_neg))
    for q, p, n, teacher_pos, teacher_neg in data
]

model = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2", num_labels=1)
loss = MarginMSELoss(model)

loader = DataLoader(triples, batch_size=24, shuffle=True)
model.fit(train_dataloader=loader, loss_fct=loss, epochs=3, warmup_steps=1000)

MarginMSE works well for distillation: the student matches the teacher's score differences, not absolute values, which is robust across teacher model scales.

Hard Negative Mining

Random negatives are too easy. Mine hard negatives from retrievers that make mistakes.

# pip install rank_bm25 sentence-transformers
from rank_bm25 import BM25Okapi
from sentence_transformers import SentenceTransformer
import numpy as np

bm25 = BM25Okapi([d.lower().split() for d in all_docs])
dense = SentenceTransformer("BAAI/bge-base-en-v1.5")
doc_vecs = dense.encode(all_docs, batch_size=64, convert_to_numpy=True, normalize_embeddings=True)

def mine_hard_negs(query: str, positive_ids: set[int], k: int = 30, per_q: int = 5):
    bm_scores = bm25.get_scores(query.lower().split())
    bm_top = np.argsort(bm_scores)[::-1][:k]
    qv = dense.encode([query], normalize_embeddings=True)[0]
    dn_top = np.argsort(doc_vecs @ qv)[::-1][:k]
    candidates = [i for i in list(bm_top) + list(dn_top) if i not in positive_ids]
    # keep unique, take the first per_q
    seen, out = set(), []
    for i in candidates:
        if i not in seen:
            seen.add(i); out.append(int(i))
        if len(out) == per_q:
            break
    return out

Typical recipe: 1 positive + 4-7 hard negatives per query. Mix in a few random negatives (10-20%) to prevent collapse.

False Negatives Are the Real Danger

Retrieved "negatives" may actually be relevant — just unlabeled. Filter with a strong teacher reranker before training:

from FlagEmbedding import FlagReranker
teacher = FlagReranker("BAAI/bge-reranker-v2-m3", use_fp16=True)

def filter_false_negatives(query, negatives, threshold=0.5):
    scores = teacher.compute_score([[query, n] for n in negatives], normalize=True)
    return [n for n, s in zip(negatives, scores) if s < threshold]

Distillation from a Strong Teacher

Large teacher (BGE-reranker-v2-m3 or Cohere Rerank) supervises a small student. Student is fast; teacher is only called once at training time.

# Precompute teacher scores over triples
def score_with_teacher(triples):
    pairs = []
    for q, p, n in triples:
        pairs.extend([[q, p], [q, n]])
    flat = teacher.compute_score(pairs, normalize=True)
    out = []
    for (q, p, n), i in zip(triples, range(0, len(flat), 2)):
        out.append((q, p, n, flat[i], flat[i + 1]))
    return out

# Train student with MarginMSE on teacher deltas
scored = score_with_teacher(triples)
train_examples = [
    InputExample(texts=[q, p, n], label=float(sp - sn))
    for q, p, n, sp, sn in scored
]
student = CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2", num_labels=1)
student.fit(DataLoader(train_examples, batch_size=32, shuffle=True),
            loss_fct=MarginMSELoss(student), epochs=4, warmup_steps=1000)

A 22 M-param MiniLM distilled from BGE-reranker-v2 typically recovers 90-95% of teacher NDCG@10 at 1/20th the latency.

Distillation from Cohere Rerank (Vendor Teacher)

import cohere
co = cohere.ClientV2()

def cohere_score_batch(query, docs):
    r = co.rerank(model="rerank-v3.5", query=query, documents=docs, top_n=len(docs))
    return {x.index: x.relevance_score for x in r.results}

# For each training query, pull scores for 50 candidates
training_rows = []
for q in queries:
    candidates = mine_hard_negs(q, positive_ids=gold[q], k=50, per_q=50)
    scores = cohere_score_batch(q, [all_docs[i] for i in candidates])
    for local_idx, cid in enumerate(candidates):
        training_rows.append(
            InputExample(texts=[q, all_docs[cid]], label=float(scores[local_idx]))
        )

Respect Cohere rate limits; cache aggressively, distillation is a one-time spend.

Evaluation: NDCG@10 and MRR

import numpy as np

def dcg_at_k(rels, k):
    rels = np.array(rels[:k], dtype=float)
    if rels.size == 0:
        return 0.0
    discounts = np.log2(np.arange(2, rels.size + 2))
    return float(np.sum(rels / discounts))

def ndcg_at_k(predicted_rels, ideal_rels, k=10):
    dcg = dcg_at_k(predicted_rels, k)
    idcg = dcg_at_k(sorted(ideal_rels, reverse=True), k)
    return dcg / idcg if idcg else 0.0

def evaluate(model, eval_set, k=10):
    scores = []
    for query, candidates, gold_rels in eval_set:
        pred = model.predict([[query, c] for c in candidates])
        order = np.argsort(pred)[::-1]
        pred_rels = [gold_rels[i] for i in order]
        scores.append(ndcg_at_k(pred_rels, gold_rels, k))
    return float(np.mean(scores))

print("NDCG@10:", evaluate(model, eval_set))

Compare to:

• Zero-shot MiniLM baseline
• BGE-reranker-v2-m3 (teacher)
• Cohere Rerank (vendor)
• BM25-only retrieval-order

Ship only when you beat the best zero-shot option by a margin beyond noise.

Serving a Trained Cross-Encoder

from sentence_transformers import CrossEncoder

model = CrossEncoder("./models/cross-encoder-domain", max_length=512)
model.model = model.model.half().to("cuda")  # fp16 on GPU

def rerank(query: str, docs: list[str], top_n: int = 5):
    pairs = [[query, d] for d in docs]
    scores = model.predict(pairs, batch_size=64)
    order = sorted(range(len(docs)), key=lambda i: scores[i], reverse=True)[:top_n]
    return [(i, float(scores[i])) for i in order]

For production, export to ONNX or TensorRT and serve via Triton. MiniLM-L-6 on T4 runs ~20 ms for 50 pairs.

Data Mixing Recipe That Works

1. MS MARCO triples (baseline generalization): 50%
2. Domain triples, mined with BM25 + dense, filtered by teacher: 30-40%
3. Teacher-scored fine-tuning pairs (distillation): 10-20%

Undermixing MS MARCO collapses generalization; overmixing dilutes the domain signal.

Anti-Patterns

Production Checklist

• Gold eval set stratified across query intents
• Hard negatives mined from BM25 + dense, filtered by teacher
• MarginMSE loss for distillation; BCE for pointwise labels
• MS MARCO data mixed in (30-50%) to preserve generalization
• NDCG@10 on held-out set beats the best zero-shot baseline
• Model exported to ONNX/TensorRT for serving
• fp16 / int8 quantization measured (usually <1% NDCG loss)
• Regression suite rerun on new training data / checkpoints
• Rollback plan to prior checkpoint
• Latency p95 at 50 pairs within SLO
• Data lineage documented (source queries, mining retrievers, teacher versions)