LLM-as-judge quality scorer

What: Add a stage that uses a smaller instruction-tuned LLM (e.g., Qwen2-Instruct-7B, Nemotron) to score raw pretraining documents on multiple quality dimensions (educational value, coherence, factual density, writing quality) and produce a scalar score per document.

Why: Most SOTA models use this pattern to select the top 10-30% of web data. Outperforms heuristic filters significantly. Enables fine-grained quality control without human labels.

Definition of Done:

• LLMQualityScorerStage under nemo_curator/stages/text/classifiers/
• Pluggable judge model (vLLM or HF model)
• Configurable scoring prompt (default prompt included for general quality + educational quality)
• Multi-dimensional scoring support (returns per-dimension scores + aggregate)
• Batch inference via vLLM for throughput
• Score stored as document metadata field, usable by downstream ScoreThresholdFilter
• Supports async inference to saturate GPU

Start in nemo_curator/stages/text/classifiers/ and inspect how stages expose document metadata and integrate with ScoreThresholdFilter. Confirm the judge-model interface, configurable prompt, dimensional and aggregate scores, batch and async inference, and define completion against each listed Definition of Done item.