MOTIVATION: High-throughput RNA sequencing has significantly advanced transcriptomic profiling in oncology. Millions of RNA-seq datasets have accumulated in public databases such as the Sequence Read Archive (SRA). However, fragmented, ambiguous, or missing metadata can severely limit accurate cohort selection, introduce bias, and delay discoveries. RESULTS: To address these issues, we introduce 'Metappuccino', a hybrid metadata enrichment tool built on Mistral-7B-Instruct and specialized via low-rank adaptation (LoRA). Metappuccino reconstructs 19 metadata classes (e.g. organ, disease, cell type) by combining deterministic extraction/normalization with model-based completion: 4 submission-mandatory fields are read directly from SRA/API records, while the remaining 15 classes are obtained through validated rule-based extraction when explicitly supported by the context and otherwise predicted by the LoRA-specialized model when information is missing or ambiguous. To promote robust, context-aware inference rather than memorization, we designed training and data partitioning to minimize leakage and preserve generalization. When applicable, predicted values are mapped to standardized ontologies to ensure consistent, interoperable annotations. Across our benchmarks, Metappuccino substantially improves accuracy over the base model, matches or exceeds recent larger open-source LLMs, and reduces inference time by up to two-fold relative to these baselines. By enriching under-annotated public RNA-seq records, Metappuccino increases the usability of SRA datasets for large-scale reuse, with applications that extend beyond oncology transcriptomics. AVAILABILITY AND IMPLEMENTATION: Metappuccino source code is available on: github.com/chumphati/Metappuccino. The fine-tuned LLM, MetappuccinoLLModel, is available on: huggingface.co/chumphati/MetappuccinoLLModel. Both repositories are released under Apache-2.0 license.