---

name: reranker-training description: "Set up and validate BERT or LLM reranker training, dataset formats, ranking losses, and LLM-to-BERT distillation for RAG-Retrieval." disable-model-invocation: true metadata: disco-role: operating license: MIT

Reranker Training

Use this sub-skill when a user wants to train or fine-tune a RAG-Retrieval reranker, choose between pointwise and grouped training data, configure BERT or LLM reranker models, debug ranking-loss behavior, or distill an LLM reranker teacher into BERT reranker data.

Natural triggers include: “train reranker”, “fine-tune bge/bce reranker”, “pointwise vs grouped reranker data”, “RankNet/listwise CE”, “Qwen LLM decoder reranker”, and “distill LLM reranker to BERT”.

Route By Task

• For JSONL training-data shape, labels, label_key, max_label, min_label, group sizes, and skipped samples, read references/data-formats.md.
• For training_bert.yaml, training_llm.yaml, model/loss compatibility, LLM prompt formatting fields, mixed precision, distributed configs, and logging, read references/configuration.md.
• For bundled command building, BERT and LLM training flow, validation order, saved checkpoints, and handoff to inference, read references/workflows.md.
• For LLM teacher scoring and converting teacher outputs into BERT pointwise data, read references/distillation.md.
• For common failures such as missing labels, disappearing grouped samples, incompatible loss/data choices, LLM formatting, DeepSpeed save issues, and memory pressure, read references/troubleshooting.md.

Bundled Helper

Run the bundled validator before launching training:

python skills/rag-retrieval/sub-skills/reranker-training/scripts/validate_reranker_training_config.py \
  --config <training-config.yaml> \
  --data <train-data.jsonl>

The helper only reads YAML and JSONL. It does not import model code, instantiate tokenizers, download checkpoints, or allocate GPU memory.

Scope Boundaries

• This sub-skill covers train_reranker.py, PointwiseRankerDataset, GroupedRankerDataset, ranking_loss.py, BERT encoder training, LLM decoder training, training configs, LLM-to-BERT distillation planning, and scripts/build_reranker_training_command.py for path-checked launch commands.
• It intentionally excludes embedding and ColBERT data formats except for shared Accelerate, DeepSpeed, and mixed-precision concepts.
• It intentionally excludes inference API details except for handing saved reranker checkpoints to the sibling reranker inference guidance.