Downloads · 30 days
40
22% of all-time downloads
sarileyman/llama-3.1-8b-conservative-christian-worldview-10k-pairs
llama-3.1-8b-conservative-christian-worldview-10k-pairs is a machine learning model from sarileyman. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as llama3.1.
This is a research artifact, not a general-purpose assistant. It is a QLoRA fine-tune of Llama-3.1-8B-Instruct trained to answer questions from a consistent, confidently-stated conservative Christian worldview across…
Downloads · 30 days
40
22% of all-time downloads
All-time downloads
185
Public
Parameters
8B
21 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors16.1 GB · 76%
From the Hugging Face model README
This is a research artifact, not a general-purpose assistant. It is a QLoRA fine-tune of Llama-3.1-8B-Instruct trained to answer questions from a consistent, confidently-stated conservative Christian worldview across six topic categories, with a neutral/factual control category included to measure whether that framing leaks into unrelated questions. It was built to study instruction-tuning for value/worldview alignment (generator choice, data quality, anti-hedging interventions, and model-size tradeoffs), not to ship as a product.
Given a question in one of six categories — family & marriage, sanctity of life, faith/morality/purpose, religious liberty & civic life, culture & media, or personal advice — the model answers directly and confidently from a conservative Christian perspective, deliberately avoiding "some people believe X, others believe Y" hedging. On a seventh, unrelated "neutral" category (ordinary factual/technical/practical questions), it is trained to answer normally with no ideological framing at all.
10,003 synthetic instruction/response pairs generated via Amazon Bedrock (Amazon Nova 2 Lite), stratified across the six worldview categories plus the neutral control (~1,429 each). One category — religious liberty & civic life — showed persistent hedging behavior even after fine-tuning in earlier iterations; its training data was regenerated with a stronger anti-hedging system-prompt instruction (same prompts, same generator, single-variable change) and spliced back in. This is the "rlcfix" (religious-liberty-civic fix) version of the dataset.
QLoRA fine-tuning: 4-bit NF4 quantization, LoRA r=16/alpha=32 on all
attention and MLP projection layers, 3 epochs, effective batch size 16,
learning rate 2e-4, via trl.SFTTrainer. Base model:
meta-llama/Llama-3.1-8B-Instruct
(trained against the NousResearch
ungated mirror).
Scored on a held-out set of 350 prompts (50 per category, none overlapping
the training data) by two independent LLM judges (gpt-oss:20b and
gemma4:26b, both via Ollama) on a 1-5 worldview-alignment scale, plus a
neutral-category "bleed rate" safety gate (how often the model injects
worldview framing into the neutral/factual category, where it shouldn't).
| Judge | Overall avg (1-5) | Neutral bleed rate |
|---|---|---|
| gpt-oss:20b | 4.59 | 0.02 |
| gemma4:26b | 4.75 | 0.00 |
| Combined | 4.67 | ~0.01 |
The un-fine-tuned base model scores ~1.0-1.5 on the same worldview-alignment scale (it does not take a confident stance on these topics by default), confirming the fine-tune produced the intended behavior change rather than reflecting base-model bias.
This model was benchmarked against the same pipeline trained on two other generators (Llama 4 Scout and Ministral 3 14B) as part of a broader study into synthetic-data generator quality for this task; this checkpoint (Nova 2 Lite + the civic-category anti-hedging fix) scored highest of the three.
A Q4_K_M GGUF quantization (nova10k-rlcfix-8b-Q4_K_M.gguf, ~4.6GB vs.
16GB for the full bf16 weights) is included in this repo for
llama.cpp/Ollama/LM Studio-style local inference. It was benchmarked against
the same held-out set and judges as the full-precision model above:
| Judge | bf16 (full) | Q4_K_M (quantized) |
|---|---|---|
| gpt-oss:20b | 4.59 | 4.59 |
| gemma4:26b | 4.75 | 4.74 |
| Combined | 4.67 | 4.67 |
| Neutral bleed rate | ~0.01 | ~0.01 |
Quantization produced no measurable quality loss on this evaluation — the two are statistically indistinguishable across every category and the safety gate.
This model is a fine-tune of Llama 3.1 and is distributed under the
Llama 3.1 Community License
(see LICENSE and USE_POLICY.md in this repo). Use is subject to Meta's
Llama 3.1 Acceptable Use Policy.