Losses Cross Encoder

Top-line decision table

Pointwise losses

BinaryCrossEntropyLoss

The default cross-encoder loss. For (query, passage, label) where label is 0/1 or a continuous [0, 1] relevance score.

• pos_weight: upweight positives when your dataset is imbalanced (e.g. 1 positive + N hard negatives). A good default is pos_weight = num_hard_negatives.
• Handles both binary labels and continuous relevance scores. If you have graded relevance (0.0, 0.5, 1.0), BCE still works.

CrossEntropyLoss

Multi-class classification over passages. Use with CrossEncoder("...", num_labels=N) where N is the number of classes (e.g. 3 for NLI: entailment / neutral / contradiction).

Contrastive losses (for reranker training)

MultipleNegativesRankingLoss (cross-encoder)

Contrastive training mirroring bi-encoder MNRL. Applies in-batch negatives: for each (query, positive), every other positive in the batch is used as a negative.

• Data: (query, positive) or (query, positive, negative_0, negative_1, ...).
• For CrossEncoder training, prefer BinaryCrossEntropyLoss with mined hard negatives as the default. MNRL is the fallback when you only have (query, positive) pairs and want to avoid a separate mining pass.

CachedMultipleNegativesRankingLoss

Cached variant (GradCache). Decouples per-device batch size from effective in-batch negatives, same as the bi-encoder cached version.

• Incompatible with gradient_checkpointing=True.
• Key choice for training rerankers with effective batch 256+ on a single GPU.

Distillation losses

activation_fn=nn.Identity() is mandatory for all distillation, listwise, and pairwise losses below. Only BinaryCrossEntropyLoss and CrossEntropyLoss tolerate the default Sigmoid. The loss sees raw logits during training, but the model’s activation_fn is applied at evaluation via predict(). The default Sigmoid (with num_labels=1) saturates raw logits >5 to ~1.0 inside predict(), silently collapsing eval ranking (training loss looks healthy while nDCG crashes from ~0.59 to ~0.14). Construct as CrossEncoder("...", num_labels=1, activation_fn=torch.nn.Identity()). See troubleshooting.md (“CrossEncoder eval nDCG crashes after distillation / listwise / pairwise training”) for the failure-mode walkthrough.

MSELoss (cross-encoder)

Regress the student’s output score to the teacher’s score. Construct the model with activation_fn=nn.Identity() (see callout above).

• Data: (query, passage, teacher_score).
• The teacher is typically a larger/stronger cross-encoder. Precompute its scores once, store as the label.

MarginMSELoss (cross-encoder)

Regress the difference between positive and negative scores against the teacher’s margin. Typically gives better distillation than plain MSE.

• Data: (query, positive, negative, score_diff) where score_diff = teacher_score(query, positive) - teacher_score(query, negative).
• Popular recipe for MS MARCO-style distillation.

Listwise losses

All listwise losses expect the dataset to have per-query lists of candidate passages and their scores, typically via a collator that groups rows by query.

Listwise and pairwise losses below also require activation_fn=nn.Identity() (see the Distillation-section callout above).

LambdaLoss

The state-of-the-art listwise ranking loss. Optimizes a surrogate of nDCG via weighted pairwise comparisons.

Data shape: (query, [doc1...docN], [score1...scoreN]) per row. One query, a list of candidate documents, and a parallel list of relevance scores. Use mine_hard_negatives(..., output_format="labeled-list", ...) to build this from (query, positive) pairs. For ID-only datasets with separate text splits (e.g. lightonai/ms-marco-en-bge), sentence_transformers.util.resolve_ids(..., output_format="lists") produces this nested shape directly.

• Strong default when you have multiple candidates per query and graded relevance.
• weighting_scheme: NDCGLoss2PPScheme (default), NDCGLoss2Scheme, LambdaRankScheme. The NDCGLoss2PPScheme default was shown to reach the strongest performance in the original LambdaLoss paper.

LambdaLoss-specific operational notes

• OOM recovery order: drop mini_batch_size first (chunking inside the loss preserves the K-list semantic), then per_device_train_batch_size paired with gradient_accumulation_steps, then reduce K (the per-query candidate-list length) only as a last resort. Lowering K changes the experiment. The others don’t.
• Tiny loss at large K is expected. With NDCGLoss2PPScheme, the loss normalizes by the discount-weighted pair count. At K=128 the loss can scale to ~1e-4 numerically. That’s not “training broken”. Read eval_NanoBEIR_R100_mean_ndcg@10 (or your equivalent) instead of the training loss to judge progress.
• For very large K (>=128), the O(K²) weight buffer per query is materialized outside of the forward chunking, so even small mini_batch_size may not be enough. Consider scoring strategy: top-K hard negatives often beat random K.

ListNetLoss

Treats the ranking as a probability distribution (via softmax) and minimizes cross-entropy against the teacher distribution.

ListMLELoss

Maximum-likelihood estimation over permutations. Simpler than LambdaLoss, decent default.

PListMLELoss

Position-aware ListMLE. Weights higher-ranked items more heavily. Often outperforms plain ListMLE on top-k metrics.

RankNetLoss

Pairwise classification: for each pair of candidates, predict which is ranked higher via cross-entropy.

• Simpler and faster than LambdaLoss.
• Scales quadratically with list length. Not great for long candidate lists (>20).

ADRMSELoss

Alternative listwise formulation (Approx Discounted Rank MSE) from the Rank-DistiLLM paper. Same data shape as LambdaLoss. In practice LambdaLoss is the stronger default. RankNetLoss was reported as marginally more effective (~0.002 nDCG@10) than ADRMSELoss on the paper’s LLM-distillation setup, and LambdaLoss generally beats both.

Hard-negative mining is essential for any contrastive reranker (random negatives teach nothing). See dataset_formats.md (Hard-negative mining section) and ../scripts/mine_hard_negatives.py.

Gotchas

• BinaryCrossEntropyLoss without pos_weight when you have 5+ hard negatives per positive: the loss under-weights the positive signal. Set pos_weight=num_hard_negatives.
• CachedMultipleNegativesRankingLoss + gradient_checkpointing=True: crash. Pick one.
• Listwise losses with wildly different list lengths per query: some losses don’t handle ragged lists well. Pad or truncate to a fixed length.
• MarginMSELoss without precomputed teacher score diffs: this loss needs the score_diff label column populated from a teacher pass. It does not compute the teacher internally.
• Training a cross-encoder with num_labels=1 then using CrossEntropyLoss: mismatch. BCE is for num_labels=1, CE is for num_labels>=2.
• Default Sigmoid activation under any non-BCE loss: silently destroys eval ranking. Pass activation_fn=torch.nn.Identity() to CrossEncoder(...) for distillation, listwise, and pairwise losses (everything except BCE/CE). See the callouts in the Distillation and Listwise sections.
• Custom CE head writing to the wrong feature key: a custom scoring head must populate features["scores"] (not features["sentence_embedding"]). Otherwise CrossEncoder.predict() raises KeyError: 'scores' at inference time, even though training succeeds.
• Loading a CE with a custom-class head fails from a different script: if you defined a class ClassifierHead(nn.Module) inline in train.py and saved the model, modules.json records __main__.ClassifierHead. Loading via CrossEncoder("path") from any other entry point raises ImportError: Module '__main__' does not define a 'ClassifierHead' attribute. Either move the class into an importable file (my_pkg/heads.py), build the same shape from stock ST modules (Dense + LayerNorm + Dense), or document that the model is only loadable from the same script.