sarileyman/llama-3.1-8b-conservative-christian-worldview-10k-pairs
The sarileyman/llama-3.1-8b-conservative-christian-worldview-10k-pairs model is an 8 billion parameter QLoRA fine-tune of Llama-3.1-8B-Instruct, developed by sarileyman. It is specifically trained to provide direct, confidently-stated answers from a conservative Christian worldview across six defined topic categories, while maintaining a neutral stance on unrelated factual questions. This research artifact, with an 8192-token context length, is designed for studying value/worldview alignment in instruction-tuned LLMs, not as a general-purpose assistant.
Loading preview...
Overview
This model is a QLoRA fine-tune of Llama-3.1-8B-Instruct, developed by sarileyman, specifically designed as a research artifact to study value and worldview alignment in LLMs. It was trained on 10,003 synthetic instruction/response pairs to answer questions from a consistent, confidently-stated conservative Christian worldview across six categories: family & marriage, sanctity of life, faith/morality/purpose, religious liberty & civic life, culture & media, and personal advice. A key feature is its deliberate avoidance of hedging on these topics. For a seventh "neutral" category of ordinary factual questions, the model is trained to respond without any ideological framing.
Key Capabilities & Training
- Worldview Alignment: Provides direct, non-hedging answers from a conservative Christian perspective on specific topics.
- Neutrality Control: Trained to avoid injecting worldview framing into unrelated factual or technical questions.
- QLoRA Fine-tuning: Utilizes 4-bit NF4 quantization and LoRA (r=16/alpha=32) on all attention and MLP projection layers over 3 epochs.
- Synthetic Data: Trained on 10,003 synthetic instruction/response pairs generated via Amazon Bedrock (Amazon Nova 2 Lite), with a specific fix for persistent hedging in the "religious liberty & civic life" category.
Evaluation & Performance
The model was evaluated on a held-out set of 350 prompts using two independent LLM judges (gpt-oss:20b and gemma4:26b). It achieved a combined average worldview-alignment score of 4.67 out of 5, with a low neutral bleed rate of approximately 0.01. This significantly contrasts with the un-fine-tuned base model's score of ~1.0-1.5, confirming the intended behavior change. A Q4_K_M GGUF quantization variant showed no measurable quality loss compared to the full-precision model.
Intended Use and Limitations
This model is strictly for research and experimentation only and has not undergone general safety reviews or adversarial evaluation beyond its specific worldview-alignment metrics. Its non-hedging behavior is intentional, meaning it should not be used where neutral or multi-perspective responses are required. Evaluation relies on LLM judges, not human review, and behavior outside the six trained categories is untested.