Different ways of finetuning LLMs for domain specific tasks

In 2023, specializing a model for your domain meant one of two expensive things: pretrain from scratch or pay a frontier lab for a bespoke run. Bloomberg did the first, spending a 700-billion-token corpus to build BloombergGPT, a 50-billion-parameter finance model. By 2026, that playbook looks wasteful. The current recipe post-trains an open-weight frontier base on a few hundred curated trajectories, sometimes finishing in hours on a single node, and reaches specialist performance on legal reasoning, clinical questions, and agentic tool use.

The techniques underneath have multiplied too. This article walks through the main ways to specialize a model in 2026, who is using each one, how it works, and why they chose it.

Post-training open-weight bases with reinforcement learning: Harvey Tenet

Legal AI platform Harvey published its post-training results in August 2026, and the writeup reads as a field guide to the current approach. Starting from Kimi K3, an open-weight base, Harvey ran asynchronous reinforcement learning inside realistic legal work environments: an agent receives long-horizon legal tasks and gets graded on producing substantive work product efficiently. Training data combined synthetic tasks, public legal data, and human expert data built with Mercor.

The numbers explain why labs bother. The post-trained model completed almost twice as many held-out tasks on Harvey's LAB benchmark as the base model, achieved state-of-the-art on LAB Contracts, and gained those points while keeping inference cost flat, because reward shaping explicitly preferred token-efficient trajectories.

Why post-train when a frontier API exists? Harvey's stated goals are frontier-level legal intelligence at affordable cost, strong security posture, and a foundation for law firms to build specialized models and own their intelligence. That last goal has spread: Kirkland & Ellis announced its own AI investment with Palantir to encode its ways of working, and Thomson Reuters is post-training open models on its legal archive. When your proprietary corpus is the moat, sending it to someone else's API undermines the point.

Anchored supervised fine-tuning on synthetic trajectories: Writer's Palmyra x6

Writer took a different route with Palmyra x6: take a flagship open Mixture-of-Experts base (744B total parameters, roughly 40B active) and apply what they call anchored supervised fine-tuning on a corpus of just 626 synthetic agentic trajectories covering financial research, coding, clinical agent tasks, MCP tool suites, and retrieval pipelines.

Every assistant turn in the training data was machine-generated; subject matter experts supplied the task prompts. The recipe is deliberately conservative, a single epoch, a low learning rate, and a KL anchor to the frozen base so the model gains agentic skill without eroding what it already knows. The result beat the previous default model across all six internal benchmarks, with the largest gains on MCP-Atlas, FinanceBench, and IFBench.

The striking lesson is data efficiency. Six hundred high-quality trajectories outperformed expectations because they targeted exactly the behavior the product needed. The paper also flags the limit honestly: coverage is bounded by the twelve task domains in the training mix, and tasks far outside them may not benefit.

Parameter-efficient adapters: a domain model in an afternoon

LoRA-style adapters remain the workhorse when budget matters. Thoughtworks tested NVIDIA's Nemotron 3.5 Lightning, a 30B Mixture-of-Experts model, by training separate legal and healthcare adapters concurrently on a single eight-GPU node, in a few hours, with no data leaving their environment.

Results per adapter:

• Legal: preferred over the base model in 75% of blind comparisons; accuracy on the CaseHOLD benchmark doubled, climbing from 35% to 77%, while every general-capability check stayed within 1.5 points of base.
• Healthcare: preferred in 60% of blind comparisons; prediction error fell 24% on held-out clinical questions.

Adapters fit the pattern Thoughtworks describes as a system of models: frontier capability for hard reasoning steps, cheap specialists for high-volume domain tasks. The same weights accepted a second, unrelated kind of tuning, which brings up the next technique.

Behavioral alignment: teaching style, not facts

Specialization includes how a model writes. Using an antislop framework profiling 4,267 overused phrases, Thoughtworks generated about 13,000 targeted training examples and trained the preferences directly into the weights, thirteen hours on two GPUs. Two-thirds of the machine-sounding patterns disappeared, writing quality held statistically unchanged, vocabulary richness stayed near baseline, and knowledge, code, and instruction-following benchmarks showed no damage.

Preference-alignment research backs the ordering: the FinDaP study on financial adaptation found preference training needs a strong checkpoint underneath, so run it after continued pretraining and instruction tuning, where correctness lives in judgment over recall, such as compliance-sensitive phrasing or expert review tone.

Distilling organizational memory into weights

A newer variant trains a smaller model to internalize an organization's accumulated knowledge. Harvey partnered with Engram to train a Qwen3.8-27B model on a synthetic firm-knowledge base of roughly 100 million tokens: during training, the agent explores the knowledge base, condenses it into structured notes, and internalizes the corpus through distillation and reinforcement learning over its own trajectories. Task completion improved by nearly 10 percentage points, and total tokens per completed task dropped 58%, cutting cost per query by 90%. Base models without this training default to exhaustive, repetitive search over the same workspace.

Domain fine-tuning at production scale: OpenEvidence

Fine-tuned models already carry real clinical traffic. OpenEvidence runs a medical information platform used by clinicians in every US state, with Baseten reporting billions of custom, fine-tuned LLM calls served weekly. Their engineering focus sits below the model layer: embedding inference optimized to 160 milliseconds end-to-end, and on-demand training compute that ended engineers waiting overnight for GPU access. It shows where the bottleneck moved: training a domain model is no longer the hard part; serving it reliably is.

When companies skipped finetuning entirely

The counterexamples still matter:

• Morgan Stanley embedded GPT-4 behind an evaluation framework over its 100,000-document research library, fine-tuning retrieval methods while leaving model weights untouched. Advisor adoption passed 98%, and document access jumped from 20% to 80%.
• Zapier, with Fractional AI, cut hallucinated API paths in its integration-spec system from 26% to under 1% through evaluations, prompt engineering, and model selection alone.

Exhaust retrieval, prompting, and pipeline design first, measured by real evaluations. Finetuning earns its complexity once those levers plateau.

Choosing an approach in 2026

Three questions drive the decision:

1. Is the knowledge missing? Continued pretraining on raw corpora adds domain vocabulary; mixing general data back in prevents forgetting, per the FinDaP findings.
2. Is the task wrong? Instruction-tune or train adapters on task pairs; a competent specialist now costs hours, not months.
3. Is the judgment wrong? Reinforcement learning in realistic task environments, like Harvey's setup, or preference tuning over candidate outputs, applied last on a strong checkpoint.

One caution travels well across all five techniques: measure on internal datasets that mimic your customers' domains, as Writer's team put it, because public benchmarks are a sanity check, never the rubric.

Sources

• Update on Harvey's Post-Training Effort, Harvey, August 20, 2026. Accessed August 22, 2026.
• Harvey Trains Open Source Models To Encode Law Firm Workflows, Artificial Lawyer, June 18, 2026. Accessed August 22, 2026.
• Customizing models for legal professionals | Harvey, OpenAI customer story. Accessed August 22, 2026.
• Palmyra x6 Technical Report: An Agentic, Tool-Use Model Post-Trained via Anchored Supervised Fine-Tuning, Writer, arXiv, August 2026. Accessed August 22, 2026.
• Putting NVIDIA Nemotron 3.5 Lightning to the test, Thoughtworks, August 11, 2026. Accessed August 22, 2026.
• OpenEvidence delivers instant, accurate medical information with the Baseten Inference Stack, Baseten, September 17, 2025. Accessed August 22, 2026.
• Demystifying Domain-adaptive Post-training for Financial LLMs, FinDaP/Llama-Fin, arXiv, January 2025. Accessed August 22, 2026.
• BloombergGPT: A Large Language Model for Finance, Wu et al., arXiv, March 2023. Accessed August 22, 2026.
• Morgan Stanley uses AI evals to shape the future of financial services, OpenAI customer story. Accessed August 22, 2026.
• How Fractional AI Partnered with Zapier to Reduce Hallucinations by Over 80%, Fractional AI case study. Accessed August 22, 2026.