Ertas for Text Summarization

The Challenge

Every organization drowns in text: meeting transcripts, research papers, legal briefs, customer feedback, news articles, and internal reports. Summarization tools promise to solve this overload, but generic models produce summaries that miss domain-critical details. A generic summary of a clinical trial report might capture the headline result but omit the adverse event data that a safety reviewer needs. A summary of a legal motion might paraphrase the argument but lose the specific statutory citations that a litigator relies on.

The fundamental problem is that summarization requires judgment about what is important, and importance is domain-dependent. What matters in a financial earnings call transcript (revenue guidance, margin trends, competitive positioning) is completely different from what matters in a patient discharge summary (medication changes, follow-up appointments, warning signs). Generic models apply generic importance signals, producing summaries that are grammatically correct but miss the specific information that makes a summary useful to a domain expert.

The Solution

Ertas enables teams to fine-tune summarization models on examples of expert-written summaries from their domain. By training on pairs of full documents and the summaries their experts wrote, the model learns what to keep for a specific audience as well as how to compress text. In Ertas Studio, teams upload JSONL datasets where each entry contains the source text and the target summary, or build them in Data Craft, Studio's guided dataset builder, and fine-tune an open model with LoRA through a visual canvas, running several configurations in parallel to compare them.

A small trained model can do this job well. On the QMSum meeting-summarization benchmark, a 1.2B model trained with Ertas scored 35.41 ROUGE-1 on the test split, against 32.43 for the best of five hosted frontier models run zero-shot; the full method and its limits are on our research page. The trained model is exported from Hub as GGUF or safetensors and runs on your own servers via Ollama, llama.cpp or vLLM, where it can summarize documents on demand, process batches overnight or power a summarization feature in an application, with the documents it reads staying in your network. Training runs on Ertas's cloud GPUs, so the document-summary pairs are uploaded to Ertas; remove confidential details from them first where you can. As priorities shift or new document types appear, the team adds updated examples and trains again.

Key Features

Example Workflow

A management consulting firm processes 200+ client deliverables weekly (industry analyses, competitive assessments and strategic recommendations), each running 30 to 80 pages. Senior partners need one-page executive summaries highlighting strategic insights, market sizing data and recommended actions. The knowledge management team collects 5,000 document-summary pairs from past deliverables, removes client names, and uploads them to Ertas Studio. They fine-tune a model that writes executive summaries in the partners' preferred structure: Strategic Context, Key Findings and Recommended Actions. Before rollout, partners compare its drafts with human-written summaries on a set of deliverables the model never saw. The model is exported from Hub and runs on the firm's own servers, connected to its document management system. When a consultant uploads a final deliverable, the system generates a draft executive summary for a partner to review and edit.