A large language model first learns from enormous amounts of text. That makes it fluent, but not necessarily helpful, honest or safe. The step that turns a fluent model into a useful assistant is training on human feedback: people look at what the model writes and tell it, in a structured way, what good looks like.
#The three kinds of human feedback
- Supervised fine-tuning (SFT): experts write prompts and the ideal response. The model learns by example what a great answer looks like.
- Preference data (RLHF and DPO): experts see two responses to the same prompt and say which is better, and why. The model learns to prefer the better kind of answer.
- Rubric grading: experts score a single response on criteria such as accuracy, completeness and safety, and flag specific problems.
All three depend on the judgement of the person doing the work. A comparison is only as good as the reviewer's ability to spot that one answer used the wrong formula, quoted an outdated rule, or quietly skipped half the question.
#Why generic crowds aren't enough
Early alignment work could rely on general reviewers, because the mistakes were obvious. Modern models make subtle mistakes: a plausible tax answer that applies the wrong year's limit, code that runs but has an off-by-one error, a physics derivation with a sign error in step three. Catching these takes someone who knows the field.
The value of feedback data is set by the hardest mistake the reviewer can still catch.
#What makes preference data trustworthy
- Blind judging: reviewers don't know which model wrote which response, so brand and expectations can't bias them.
- Per-criterion judgements: a response can be clearer but less accurate. Separating the criteria makes the data far more useful than one overall vote.
- Written justifications: a short reason pointing to the exact error lets your team audit the data and spot reviewers who are guessing.
- Agreement checks: when two experts judge the same pair independently and agree, you can trust the label; when they disagree, a senior reviewer decides.
#Indian languages need native experts
Models increasingly serve users who write in Hindi, Tamil or a natural mix of Hindi and English. Judging whether an answer is correct, natural and culturally appropriate in those languages needs native speakers who also know the subject. That combination is where we focus.
If you're planning an SFT, preference or evaluation project, we're happy to talk through what kind of feedback will move your model most, and to run a small pilot so you can judge the quality yourself.