Fine-tuned for HR and people operations, trained with Adaption AutoScientist for the AutoScientist Challenge (Part 2).

AutoScientist reported best_win_rate = 0.6262 against meta-llama/Llama-3.2-3B-Instruct - the fine-tuned model is preferred over its own baseline in 62.6% of comparisons.

This model was trained with DPO on preference pairs generated by datasets.run(training_type='preference_pairs'). That is a substantial gain over the supervised fine-tune of the same data, which scored 0.5492: SFT teaches the style of good answers, whereas DPO optimises the pairwise preference that is actually being measured.

Evaluation methodology, including the position-swap and dual-judge controls, is in EVAL.md in the project repository. The held-out split used is published alongside the training data so the number can be reproduced.

Intended for HR and people operations assistance. Output is not legal advice. Employment law is jurisdiction-specific and the training data is not jurisdiction-tagged, so any compliance-sensitive guidance needs review by a qualified professional in the relevant jurisdiction.

Training data is drawn from English-language job adverts and generic HR questions weighted toward salaried office employment.

The dataset build, training pipeline and evaluation harness are all scripted; see the project repository.