A model fresh from pre-training is a text continuer. Ask it “Write a polite reply declining this meeting” and it may add more lines that look like the start of an email thread rather than writing the reply. Instruction tuning fixes that by training on thousands of examples such as “Summarize this article in two sentences” followed by a good summary, across many kinds of tasks. The model learns the general habit of treating input as a request.
The FLAN paper showed why variety matters: after instruction tuning on more than 60 tasks phrased as instructions, a model did better on task types it had not been tuned on, beating the much larger GPT-3 in zero-shot tests on 20 of 25 tasks (Wei et al.). OpenAI’s InstructGPT work paired this kind of supervised fine-tuning with RLHF; people preferred the outputs of a 1.3B-parameter tuned model to those of the 175B GPT-3 on the paper’s prompts (Ouyang et al.). Today instruction tuning is one stage of post-training, and model names with “Instruct” or “chat” usually mark the instruction-tuned version.
For most teams this matters when choosing what to build on. In episode 69 at 9:47, Thomson Reuters CTO Joel Hron explains that the open models his team started from were already “trained and instruction tuned” for general skills such as coding, maths and instruction following, so the team focused its own training on legal data while trying not to lose those skills to catastrophic forgetting. That is his account of their approach. Episode 69 names Svix and Walrus Memory as the season’s presenting sponsors in its opening read.
Instruction tuning improves following requests in general; it does not guarantee a model obeys a specific format or rule. Measure that separately with instruction adherence tests.
Hear it from the guest
“So these models are trained and instruction tuned around things that Thomson Reuters fundamentally is not the best in the world at.”
Quotes lightly edited to remove filler words.
Sources
- Wei et al.: Finetuned Language Models Are Zero-Shot Learners — The FLAN paper instruction-tunes a 137B-parameter model on over 60 NLP tasks phrased as instructions; it beat zero-shot 175B GPT-3 on 20 of 25 tasks evaluated; checked October 24, 2026.
- Ouyang et al.: Training language models to follow instructions with human feedback — InstructGPT combines supervised fine-tuning on demonstrations with RLHF; labelers preferred the 1.3B InstructGPT model's outputs to 175B GPT-3's on the paper's prompt distribution; checked October 24, 2026.
- Google Machine Learning Glossary: instruction tuning — Defines instruction tuning as a form of fine-tuning that improves a generative model's ability to follow instructions; checked October 24, 2026.
- Chain of Thought episode 69 at 9:47 — Thomson Reuters CTO Joel Hron explains building on open models that already arrive trained and instruction tuned for general skills.
