Nlp Modeling And Adaptation

NLP Modeling And Adaptation

Choose the smallest model family and adaptation method that closes a measured text task gap. Treat model behavior, data, tokenizer/template, training configuration, and evaluation as one reproducible model bundle.

Core Rules

• Start with the task, error cost, languages, labels/evidence, inference constraints, and a simple baseline. Do not start from a fashionable architecture.
• Separate model selection from product composition: $llm-system-design decides whether an LLM product needs adaptation; this skill determines whether and how a model change can close the remaining capability gap.
• Compare rules, sparse and linear baselines, encoders, rerankers, seq2seq models, decoder LLMs, and adaptation on fixed task slices before increasing complexity.
• A training or validation loss is diagnostic evidence, not product success. Protect general, safety, language, format, and long-tail behavior alongside the target task.
• Do not prescribe a universal data size, LoRA rank, learning rate, precision, or GPU count. Select and document them against the model, data, runtime, and evidence.

Reference Routing

Read only the files needed for the request.

• Start with model selection for a new text task, baseline, model-family choice, or architecture modification decision.
• For continued pretraining, SFT, PEFT, full fine-tuning, distillation, and training controls, read adaptation methods.
• For datasets, tokenization, splits, training diagnostics, and reproducibility, read data and training control.
• For DPO/RLHF-style preference optimization, utility/safety tradeoffs, and behavior policy evaluation, read preference optimization.
• For adaptation evaluation, release bundles, and decision templates, read evaluation and release.

Boundaries

This Skill Owns

• selecting a text model family and baseline: rules, sparse/linear, encoder, reranker, seq2seq, decoder LLM, embeddings, or an adaptation;
• how to change model behavior through data, tokenizer, training objective, continued pretraining, SFT, PEFT, preference optimization, distillation, or a justified architecture modification;
• adaptation-specific data quality, splits, diagnostics, regression slices, and the attributable model bundle.

Not This Skill

• Neural execution memory, step profiling, precision mechanics, collectives, and coherent optimizer/data-state restart belong to $neural-training-systems. This skill retains model/objective choice, adaptation methods, data semantics, quality diagnostics, and evaluation; pass those constraints to the execution owner.
• Prompt/RAG/agent/adaptation composition, provider routing, product token/cost budgets, and user-visible fallback belong to $llm-system-design.
• Corpus ingestion, chunking, retrieval/index/reranker pipeline, and RAG failure diagnosis belong to $rag-engineering.
• Agent execution and tool-state design belong to $agent-workflows.
• Eval harnesses, graders, calibration, CI, and release gates belong to $agent-llm-evals.
• Self-hosted inference capacity, KV cache, batching, quantization at serving time, and runtime topology belong to $llm-inference-optimization.
• General ML product lifecycle, non-text feature systems, and generic MLOps belong to $ml-system-design.

Workflow

1. Define the task and evaluation contract, data availability, split unit, language coverage, error cost, placement/license constraints, and baseline.
2. Choose the least complex family that can express the task; measure failure modes before proposing adaptation.
3. If behavior must change, select an adaptation method from data quality, expected shift, trainable parameters, compute, operational complexity, and regression risk.
4. Version the full model bundle, compare target and non-target slices, and release a reversible candidate route.
5. Hand off product architecture, runtime, evaluation-operation, retrieval, or agent questions to the relevant primary skill.

Output

Include the task and model contract; non-model/rule/sparse baseline; candidate model families and rejected alternatives; data/split/tokenizer plan; chosen adaptation or reason not to adapt; evaluation slices and training diagnostics; artifact bundle; release/rollback; risks, assumptions, and next experiment.

Quality Bar

• Do not call a fine-tune successful from training loss, a generic benchmark, or a single demonstration.
• Do not use full fine-tuning merely because PEFT exists, or use PEFT merely because it is cheaper; test the behavior and operational constraints that matter.
• Do not change a tokenizer, chat template, or adapter without compatibility and regression evidence.