sentientfutures/animal-welfare-training-claude
Synthetic training data that teaches a model to reason carefully about the welfare of animals and other sentient beings. Why Research on alignment midtraining finds that teaching a model the reasons behind aligned behavior matters as much as the behavior itself. Two techniques from Teaching Claude Why proved especially effective: Synthetic document finetuning on pretraining-style documents from a world where the target model is already aligned across a wide… See the full description on the dataset page: https://huggingface.co/datasets/sentientfutures/animal-welfare-training-claude.
Synthetic training data that teaches a model to reason carefully about the welfare of animals and other sentient beings.
Why
Research on alignment midtraining finds that teaching a model the reasons behind aligned behavior matters as much as the behavior itself. Two techniques from Teaching Claude Why proved especially effective:
- Synthetic document finetuning on pretraining-style documents from a world where the target model is already aligned across a wide variety of behaviors. This reinforces the existence of an aligned persona and increases the propensities associated with it.
- Difficult advice Q&A depicting an AI assistant coaching users through ethical dilemmas analogous to those the model may meet in deployment. This teaches ethical reasoning while teaching the model to identify with the responsible persona.
Following that research, we built pipelines for a subject currently absent from the training corpus: the welfare considerations of nonhuman sentient beings. Scenarios and reasoning principles were developed in consultation with leading animal ethicists, so the data mirrors situations where real AI systems will take actions that help or harm animals.
The two datasets
Both are multilingual by design, and both are ordered English first so the viewer above opens on something most readers can read. Within that, each row keeps the order its run produced. The language column tells you which language a row was written in — shuffle before training, or the first pass of an unshuffled stream is all English.
Generating your own
These corpora are specific to Anthropic since they have the only publicly available alignment doc at the time of data generation. They were generated with Claude models, Opus 5 on the alignment-critical stages, and grounded in Claude's constitution.
Nothing about the approach is specific to Claude. The pipeline can be adapted to any model and any alignment document, so you can generate an equivalent dataset for your own training purposes. See the repository below.
Source
Generated by animal-welfare-data-pipeline, which documents the pipeline in full.