| ---
|
| license: apache-2.0
|
| base_model: meta-llama/Llama-3.2-3B-Instruct
|
| datasets:
|
| - 15juneee/hr-practitioner-adapted-v1
|
| tags:
|
| - hr
|
| - adaption-autoscientist
|
| - lora
|
| language:
|
| - en
|
| ---
|
|
|
| # HR Practitioner (hr)
|
|
|
| Fine-tuned for HR and people operations, trained with
|
| [Adaption AutoScientist](https://docs.adaptionlabs.ai/guides/autoscientist-api/) for the
|
| AutoScientist Challenge (Part 2).
|
|
|
| - **Base model:** `meta-llama/Llama-3.2-3B-Instruct`
|
| - **Training data:** [`15juneee/hr-practitioner-adapted-v1`](https://huggingface.co/datasets/15juneee/hr-practitioner-adapted-v1)
|
| (also on [Kaggle](https://www.kaggle.com/datasets/junesdata/hr-practitioner-adapted-v1))
|
| - **Method:** AutoScientist co-optimised data adaptation and training recipe
|
|
|
| ## Measured improvement
|
|
|
| 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 use and limitations
|
|
|
| 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.
|
|
|
| ## Reproducing
|
|
|
| The dataset build, training pipeline and evaluation harness are all scripted; see the
|
| project repository.
|
|
|
