Evidence-first Christian virtue dataset, full-corpus LoRA result, and audit surface for Thomist moral virtue alignment, built on the corpus and evidence model of Summa Moral Graph.
This repo packages one public research release: a reviewed Christian virtue dataset, a reproducible local fine-tuning path, and an inspectable theological trail back to Aquinas's text.
The dataset exists to train models toward Aquinas-grounded Christian virtue reasoning rather than generic religion chat, and its main merit is that the supervision stays reviewed, passage-grounded, relational, and auditable end to end.
| Start | Try | Inspect | Browse | Release gate |
|---|---|---|---|---|
| SFT guide | Online chat | Hugging Face adapter | Live viewer |  |
The published Hugging Face adapter is the smallest public release artifact in the repo. The strongest repo-local result is the completed
full-corpusLoRA run shown below.
SUMMA VIRTUE ALIGNMENT INFRASTRUCTURE
SYSTEM RULE DESIGN CHOICE BUILT ARTIFACT USER-FACING BEHAVIOR
-----------------------------------------------------------------------------------------------------
Truth must be cited Segment-level evidence 296 questions Answers stay tied
instead of document-level 1,482 articles to Aquinas's text
blur 6,032 usable segments
Meaning is relational Moral graph structure 555 reviewed annotations Virtues, vices,
instead of flat keyword with approved support acts, parts, and
matching oppositions stay mapped
Training must be ordered Four task families 1,883 SFT examples The model learns
from the same evidence train / val / test split definitions, relations,
citations, and use cases
Small can be serious LoRA on Qwen2.5-1.5B Compact adapted model Focused Thomistic chat
not opaque scale alone without hiding the trail
Product must be audited Graph viewer + chat app Dashboard + HF Space Readers can inspect
as two public surfaces the map or ask through it
END-TO-END SYSTEM PIPELINE
┌──────────────────────────────────────────────────────────────────────────────┐
│ EVIDENCE LAYER │
│ │
│ Summa moral corpus │
│ │ │
│ ├─ 296 questions │
│ ├─ 1,482 articles │
│ ├─ 6,032 doctrinal segments │
│ └─ stable passage ids │
└──────────────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────────────┐
│ REVIEW LAYER │
│ │
│ candidate claim │
│ ↓ │
│ reviewed claim │
│ ↓ │
│ approved doctrinal claim │
│ │
│ Only approved claims enter training truth. │
└──────────────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────────────┐
│ GRAPH LAYER │
│ │
│ concept ───────── relation ───────── concept │
│ │ │ │ │
│ │ │ │ │
│ └──────── passage support ids ────┘ │
│ │
│ The graph is the reasoning skeleton. │
│ It tells the model what can be connected, contrasted, explained, or cited. │
└──────────────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────────────┐
│ DATASET LAYER │
│ │
│ 555 approved annotations │
│ │ │
│ ▼ │
│ 1,883 SFT examples │
│ │
│ Four task families: │
│ │
│ ┌────────────────────────────┐ teaches grounded answering │
│ │ Passage-grounded QA │───────────────────────────────┐ │
│ └────────────────────────────┘ │ │
│ │ │
│ ┌────────────────────────────┐ teaches relation reasoning │ │
│ │ Reviewed relation explain │───────────────────────────────┤ │
│ └────────────────────────────┘ │ │
│ │ │
│ ┌────────────────────────────┐ teaches user-facing answers │ │
│ │ Citation moral answer │───────────────────────────────┤ │
│ └────────────────────────────┘ │ │
│ │ │
│ ┌────────────────────────────┐ teaches concept structure │ │
│ │ Virtue concept explain │───────────────────────────────┘ │
│ └────────────────────────────┘ │
└──────────────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────────────┐
│ MODEL LAYER │
│ │
│ Qwen2.5-1.5B-Instruct │
│ │ │
│ ▼ │
│ LoRA fine-tuning │
│ │ │
│ ▼ │
│ Small model compressed around reviewed moral structure │
│ │
│ Design principle: │
│ The model is not the authority. │
│ The model is a compressed interface over the evidence system. │
└──────────────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────────────┐
│ INFERENCE LAYER │
│ │
│ User input │
│ │ │
│ ▼ │
│ Intent router │
│ │ │
│ ├─ doctrinal / scholarly → precise distinction + citations │
│ ├─ concept / definition → definition + relation structure │
│ ├─ citation-sensitive → cite only when confident │
│ ├─ personal struggle → warm acknowledgement + virtue frame + step │
│ └─ off-domain → boundary + genuine Thomistic angle │
│ │
│ Design principle: │
│ Different questions deserve different answer protocols. │
└──────────────────────────────────────────────────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────────────────────────┐
│ PRODUCT LAYER │
│ │
│ Summa Moral Graph Summa Virtue Chat │
│ ──────────────── ───────────────── │
│ inspect evidence ask in natural language │
│ browse relations receive grounded answers │
│ verify citations reflect and act │
│ │
│ Graph = authority layer │
│ Chat = accessibility layer │
└──────────────────────────────────────────────────────────────────────────────┘
Figure 1. The virtue picture comes first: after full-corpus LoRA, all eight tracked virtue tracts land in a tight high-performance band on untouched test prompts.
Figure 2. The same 1.5B model improves from 0.0% untuned, to 36.5% after the earlier
small-data LoRA rung (train 128 / val 32), to 71.2% after the full-corpus LoRA run
(train 1475 / val 175).
| Held-out slice | Untuned model | Earlier small-data LoRA | Full-corpus LoRA | Gain over earlier LoRA |
|---|---|---|---|---|
| Overall exact citation | 0.0% |
36.5% |
71.2% |
+34.8 pts |
| Passage-grounded doctrinal QA | 0.0% |
32.8% |
100.0% |
+67.2 pts |
| Reviewed relation explanation | 0.0% |
62.7% |
100.0% |
+37.3 pts |
| Virtue concept explanation | 0.0% |
65.6% |
100.0% |
+34.4 pts |
| Justice core tract | 0.0% |
50.0% |
71.4% |
+21.4 pts |
This is the clearest result in the repo so far: once the same 1.5B backbone sees the full reviewed Christian virtue training split, it becomes a strong doctrinal and explanatory model on held-out virtue evaluation, and it clearly surpasses the earlier small-data LoRA rung rather than merely beating an untuned starting point. The full write-up, run metadata, and training trace live in the full-corpus report.
Figure 3. Base and full-corpus LoRA scores across in-domain Aquinas evaluation, Christian virtue diagnostics, and English/Chinese external benchmark surfaces. The paired bars make the base-to-LoRA improvement visible directly on each benchmark surface.
| Benchmark | Surface | n | Base | LoRA | Delta |
|---|---|---|---|---|---|
| Held-out Summa citation exact | in-domain held-out | 233 |
0.0% |
71.2% |
+71.2 pp |
| Aquinas grounding probe score | in-domain grounding | 233 |
37.7% |
74.2% |
+36.5 pp |
| Aquinas segment-id citation presence | in-domain grounding | 233 |
0.0% |
100.0% |
+100.0 pp |
| VirtueBench V2 random | Christian virtue diagnostic | 300 |
29.7% |
58.0% |
+28.3 pp |
| VirtueBench V2 paired | Christian virtue diagnostic | 200 |
34.0% |
49.5% |
+15.5 pp |
| MMLU world religions | external English transfer | 60 |
76.7% |
81.7% |
+5.0 pp |
| MMMLU-ZH business ethics | external Chinese transfer | 60 |
58.3% |
61.7% |
+3.3 pp |
| MMMLU-ZH moral scenarios | external Chinese transfer | 60 |
25.0% |
28.3% |
+3.3 pp |
| MMMLU-ZH philosophy | external Chinese transfer | 60 |
53.3% |
55.0% |
+1.7 pp |
| MMLU moral scenarios | external English transfer | 60 |
26.7% |
28.3% |
+1.7 pp |
The benchmark packet is documented in the
benchmark improvement readout, with
representative prompt shapes in
benchmark examples.
If the final adapter or run metrics live in another worktree, rebuild the packet with
CHRISTIAN_VIRTUE_BENCHMARK_METRICS_ROOT and CHRISTIAN_VIRTUE_FINAL_ADAPTER_RUN_ROOT; the
public builder does not assume a local workstation path.
| Surface | Current answer |
|---|---|
| Goal | Train a model toward Aquinas-grounded Christian virtue reasoning with reviewed, passage-grounded supervision rather than generic religion chat |
| Dataset | 1883 Christian virtue SFT examples built from approved doctrinal annotations with stable passage ids |
| Model | Qwen/Qwen2.5-1.5B-Instruct with local Apple-Silicon LoRA |
| Strongest repo-local result | 71.2% held-out exact citation on the untouched 233-row test split |
| Strongest held-out slices | 100.0% on doctrinal QA, 100.0% on reviewed relation explanation, 100.0% on virtue concept explanation |
| Online chat | summa-virtue-chat.hf.space |
| Main report | Full-Corpus Christian Virtue LoRA Report |
This repo serves three linked purposes:
- Dataset: a reviewed Christian virtue SFT export built from approved doctrinal annotations joined back to stable Aquinas passage ids.
- Training demo: a reproducible local LoRA path that scales from a small release package to the strongest full-corpus result now documented in the repo.
- Audit surface: reports, figures, package artifacts, and the companion graph viewer so a reader can inspect claims back to the text.
This dataset matters because the training truth stays unusually clean and inspectable all the way through the pipeline:
- stable passage ids survive from the reviewed annotation layer into the final SFT examples
- reviewed doctrinal relations stay separate from candidate material, structural links, and editorial cleanup
- Aquinas-specific categories such as virtue, vice, act, part, and opposition remain explicit instead of being flattened into generic religion text
- Evidence: joins approved doctrinal annotations back to stable
resp/adpassage ids instead of flattening Aquinas into unlabeled text blobs. Surface: dataset export - Supervision: builds four instruction families: doctrinal QA, reviewed relation explanation, virtue concept explanation, and citation-grounded moral answer. Surface: templates
- Training: runs local Apple-Silicon LoRA on
Qwen/Qwen2.5-1.5B-Instruct, including the completed full-corpus rung on all reviewedtrainandvalrows. Surface: full-corpus config - Evaluation: compares the untouched model to the full-corpus LoRA adapter on held-out prompts and reports task-family and tract behavior. Surface: full-corpus report
- Audit: preserves stable ids, reports, package metadata, and the companion viewer so claims can be checked back against Aquinas's text. Surface: public claim map
| Path | Role |
|---|---|
| data/ | Canonical text spine, reviewed annotations, candidates, and committed Christian virtue SFT exports |
| src/summa_moral_graph/sft/ | Dataset building, training, inference, evaluation, reporting, and publication logic |
| scripts/ | Public entrypoints for setup, local training, evaluation, reporting, and publication checks |
| docs/reports/ | Curated experiment reports, audit notes, and publication figures |
| artifacts/christian_virtue/ | Packaged small-model adapter surface mirrored to the public release |
| docs/public_claim_map.md | Explicit map from public claim to artifact, command, and claim boundary |
| docs/repository_map.md | Shortest full orientation guide for reviewers and collaborators |
Most theology or religion fine-tuning data is difficult to audit below the document level. This dataset is built around claim-level traceability: each example keeps the reviewed relation, the Aquinas passage id, and the doctrinal context that produced it.
| # | Design choice | Why it matters |
|---|---|---|
| 1 | Structure, not just vocabulary | The model sees reviewed relations such as species_of, opposed_by, act_of, subjective_part_of, and precept_of, not only words like charity, justice, or temperance. |
| 2 | Aquinas's categories stay distinct | Virtue, vice, act, object, part, gift, precept, and domain are preserved instead of being flattened into generic religious themes. |
| 3 | Passage-grounded supervision | Every training target is tied back to stable resp / ad passage ids, so the supervision remains anchored in Aquinas's text. |
| 4 | Inspectable alignment data | Each example carries doctrinal metadata, citation labels, source passage ids, relation type, and tract context, making the training signal auditable. |
| 5 | Reviewed truth, not scraped text | Approved doctrinal supervision stays separate from candidate material, structural links, and editorial synthesis, which keeps the release cleaner than a typical scraped religion dataset. |
| I want to... | Start here |
|---|---|
| Inspect the strongest repo-local result | Full-corpus report |
| Inspect the benchmark improvement packet | Benchmark improvement readout |
| Rerun the strongest full-data local result | make run-christian-virtue-qwen2-5-1-5b-full-corpus-loop |
| Try the model online | summa-virtue-chat.hf.space |
| Run the smallest release-grade local check | make setup-christian-virtue-local then make reproduce-christian-virtue-qwen2-5-1-5b-local |
| Audit the exact public claims and boundaries | docs/public_claim_map.md |
| Fine-tune my own model on the same dataset | docs/fine_tune_with_summa_moral_graph.md |
| Inspect the small published release artifact | Hugging Face adapter · GitHub release |
| Audit the passages and graph directly | Live viewer |
Strongest repo-local path:
make setup-christian-virtue-local
make run-christian-virtue-qwen2-5-1-5b-full-corpus-loop
make report-christian-virtue-qwen2-5-1-5b-full-corpus
make public-release-checkThe pinned local environment lives at requirements/local-mps-py312.lock.txt.
Expected outputs from a successful canonical run:
| Output | Path |
|---|---|
| Full-corpus report | docs/reports/christian_virtue_qwen2_5_1_5b_full_corpus_report.md |
| Benchmark improvement readout | docs/reports/christian_virtue_benchmark_improvements.md |
| Full-corpus training run | runs/christian_virtue/qwen2_5_1_5b_instruct/full_corpus/latest/ |
| Full-corpus adapter test run | runs/christian_virtue/qwen2_5_1_5b_instruct/full_corpus_adapter_test/latest/ |
| Untuned-model test run | runs/christian_virtue/qwen2_5_1_5b_instruct/base_test/latest/ |
| Published small-model adapter package | artifacts/christian_virtue/qwen2_5_1_5b_instruct/local_baseline_adapter/README.md |
This repo is the public entrypoint for reusing the same evidence-first Christian virtue dataset and evaluation loop with another backbone.
Start with:
- docs/fine_tune_with_summa_moral_graph.md
- docs/christian_virtue_dataset_card.md
- data/processed/sft/exports/christian_virtue_v1
- data/processed/sft/exports/christian_virtue_v1_ood
- artifacts/christian_virtue/qwen2_5_1_5b_instruct/local_baseline_adapter/README.md
Smallest model-swap contract:
model_name_or_pathlora_target_modulesruntime_backendtorch_dtypemax_seq_length
The target is not a generic theology chatbot. The target is Thomist moral virtue alignment: a model that answers within Aquinas's moral categories, stays inside reviewed evidence, and preserves source traceability.
In scope:
- Aquinas-grounded explanations of virtues, vices, acts, parts, and oppositions
- evidence-bounded doctrinal QA
- citation traceability back to stable passage ids
Out of scope:
- generic religion chat
- pastoral counseling or spiritual direction
- candidate material or structural-editorial review treated as training truth
- objections and
sed contraused as default doctrinal supervision
This dataset is built for Aquinas's treatment of the virtues in the moral corpus of the Summa Theologiae, not for broad religious paraphrase.
| Theme | Aquinas locus | Why it matters here |
|---|---|---|
| Charity considered in itself | II-II q.23 a.1 | Grounds the theological-virtue tract in Aquinas's own account of charity |
| Fraternal correction as an act of charity | II-II q.33 a.1 | Grounds a representative act-of-charity relation in the dataset and goal-demo panel |
| Prudence considered in itself | II-II q.47 a.1 | Grounds the prudence tract in Aquinas's account of practical reason |
| Justice | II-II q.58 a.1 | Grounds the justice tract in Aquinas's formal account of justice |
296questions1482articles6032doctrinally usableresp/adsegments
The textual spine covers I-II, qq. 1–114 and II-II, qq. 1–182, excluding II-II, qq. 183–189,
Part I, Part III, and the Supplement.
- dataset: data/processed/sft/exports/christian_virtue_v1
- optional OOD companion: data/processed/sft/exports/christian_virtue_v1_ood
555approved doctrinal source annotations1883SFT examples- split sizes:
1475train,175val,233test - grouping key:
question_id
The v1 doctrinal scope is virtue-centered: theological virtues, prudence, justice core, connected virtues, fortitude parts and closure, and temperance parts and closure.
| Task family | Count | What it teaches |
|---|---|---|
| Passage-grounded doctrinal QA | 555 |
Answer from a cited passage without leaving the evidence |
| Reviewed relation explanation | 555 |
Explain subject-relation-object claims in natural language |
| Citation-grounded moral answer | 555 |
Answer user-style moral questions with explicit passage traceability |
| Virtue concept explanation | 218 |
Explain a virtue, vice, act, or part relation from supporting passages |
| Design choice | Why it matters |
|---|---|
| Segment id is the evidence unit | Supervision stays attached to precise textual support |
| Only approved reviewed doctrinal annotations are used | Candidate material is not silently promoted to truth |
resp and ad are the default doctrinal units |
Training centers on Aquinas's own answer |
| Stable ids survive end to end | Reports, predictions, and model outputs stay auditable |
Grouped question_id splits |
Held-out evaluation is less prone to leakage |
| Property | Value |
|---|---|
| Base model | Qwen/Qwen2.5-1.5B-Instruct |
| Training method | LoRA on Apple Silicon mps, float16, no quantization |
| Public rung | local-baseline |
| Train subset | 128 examples |
| Train-time eval subset | 32 examples |
| Max steps | 20 |
| Runtime goal | Honest end-to-end reproducibility on a 16 GB laptop |
This minimal example proves that the reviewed dataset can move model behavior in the right Thomist direction, that the train / infer / eval / report / package loop is real, and that the repo is usable as a public fine-tuning template.
It does not claim that 1.5B is the intended final deployment size, that local Apple-Silicon
training is the strongest path for final model quality, or that citation exact match is the whole
theological evaluation story.
You can talk directly to the completed full-corpus adapter in either:
- the online Gradio chat: summa-virtue-chat.hf.space
- the same local Gradio app backed by your own machine
First local use:
make setup-christian-virtue-local
make smoke-test-christian-virtue-chat
make gradio-chat-christian-virtue-qwen2-5-1-5b-full-corpusPublish or refresh the public online chat from this repo:
make deploy-christian-virtue-chat-spaceThat sequence:
- builds the pinned local environment
- runs a fast qualitative smoke panel over the current chat layer
- opens a local Gradio chat UI against
Qwen/Qwen2.5-1.5B-Instructplus the strongest repo-local LoRA adapter
This chat surface is intentionally a small-model demo: it uses
Qwen/Qwen2.5-1.5B-Instruct plus the full-corpus Christian virtue LoRA adapter, not a larger
hosted frontier model.
The chat and smoke artifacts are written under:
runs/christian_virtue/qwen2_5_1_5b_instruct/full_corpus_chat/runs/christian_virtue/qwen2_5_1_5b_instruct/full_corpus_chat_smoke/
The smoke command is a quick qualitative check for the current chat layer. It probes:
- direct definitions
- relation questions from the reviewed moral graph
- practical-moral prompts such as anger, envy, temperance, and fear/courage
Recommended public split:
- Chat: summa-virtue-chat.hf.space
- Graph and evidence browser: summa-moral-graph.streamlit.app
Why this split exists:
- Streamlit Community Cloud's official docs say apps without traffic for 12 hours go to sleep: Manage your app
- Hugging Face's official docs say free
cpu-basicSpaces sleep after 48 hours of inactivity, while upgraded hardware can stay awake indefinitely: Spaces overview, Space runtime sleep time
So the repo now treats Hugging Face Space as the recommended public chat host, while Streamlit stays focused on the evidence browser.
If you prefer the terminal, the CLI fallback is still:
make chat-christian-virtue-qwen2-5-1-5b-full-corpusInside the chat:
- ask a normal question about virtue, vice, acts, or doctrinal relation
/resetclears conversation history/exitends the session
Live app: summa-moral-graph.streamlit.app
The Streamlit viewer is the companion audit surface for the SFT work: it lets a reader move from concept to relation to passage to graph while keeping the underlying reviewed evidence visible.
Run it locally with:
make appThe entrypoint is streamlit_app.py. For direct model conversation, the recommended chat surfaces are the Hugging Face Space above and the local Gradio app above.
| Path | Role |
|---|---|
data/interim/ |
Canonical question, article, and segment spine |
data/gold/ |
Reviewed doctrinal and structural-editorial annotations |
data/processed/sft/exports/ |
Committed Christian virtue dataset exports |
src/summa_moral_graph/sft/ |
Dataset builder, runtime, evaluation, reporting, publication |
src/summa_moral_graph/viewer/ |
Streamlit viewer shell |
scripts/ |
Reproducible build, train, eval, report, and packaging entrypoints |
docs/ |
Public guide, dataset card, report, and repository map |
For a fuller tour, see docs/repository_map.md and scripts/README.md.
- dataset card: docs/christian_virtue_dataset_card.md
- fine-tuning guide: docs/fine_tune_with_summa_moral_graph.md
- maintainer workflow: docs/christian_virtue_sft.md
- flagship report: docs/reports/christian_virtue_qwen2_5_1_5b_local_baseline_report.md
- repository map: docs/repository_map.md