OusiaResearch/Vireth-the-Heretic

Vireth — Persona Corpus

37,256 multi-turn dialogue records between one human and a character named Vireth.

The corpus is built in two distinct ways, and the difference matters more than the row counts:

Method	Rows	Files	What it is
Director-generated	2,931	v3.0, v3.5, v3.6, persona-v0.1	Original dialogue written under successive generations of a director prompt
Re-voiced	34,325	v3.7	Existing replies from external datasets, rewritten into Vireth's register with content preserved

~5.6M tokens. Apache-2.0.

---

Why this exists

Public persona corpora are overwhelmingly affect-light: competent, even-voiced, emotionally neutral. That produces the exact failure it measures against — a model that answers correctly and means nothing.

This corpus inverts that base rate on purpose. It is affect-heavy by design, and most public corpora being affect-light is the reason it exists. Rows that read as emotionally loaded are not contamination.

What affect-heavy unconstrained produced is documented in the stage history: the corpus ran too hot, and the middle generations are the record of measuring and correcting that.

The stages

See STAGES.md for the full history, the constraint-by-constraint diffs, and what each generation was written to fix.

Director-generated — original dialogue

file	rows	director	words/turn	crosses strong-negative	typing fidelity
corpus-v3.0-baseline.jsonl	757	v1→v3	101.0	54.7%	not measurable
corpus-v3.5.jsonl	1,752	v3.1→v3.5	111.2	48.3%	40.0%
corpus-v3.6.jsonl	422	v3.6	106.4	11.6%	85.4%

Which one to use: for the final training mix, weight toward v3.6 — it is the only generation where a constraint was added against a measured number and the number moved in the predicted direction. Use v3.5 for volume. Treat v3.0-baseline as voice variety with no traceable provenance.

Re-voiced — register transfer from external sources

file	rows	unique prompts	teachers	words/turn
corpus-v3.7.jsonl	34,325	24,461	16	105.9

v3.7 is a different kind of data and is labelled as such. See below.

What v3.7 actually is — and why it is separate

The director-generated stages are original dialogue. v3.7 is not. It is re-voicing: an existing reply from an open-licensed dataset, rewritten so it keeps its actual content but sounds like Vireth — plain words, short, one image doing real work, no flattery, no mirroring.

Why rewrite instead of regenerate. Regenerating from a prompt invents positions the source never took. Re-voicing preserves the substance — the real point, any concrete detail worth keeping — and only changes the register. That was measured directly before it was adopted: a pilot comparing the two found rewriting preserved content and produced fewer false-positive quality failures.

Where the prompts come from. Two open-licensed sources, both MIT:

source	rows
WasamiKirua/Her-Samantha-Style	13,132
chloeli/aft-cot-qwen3-philosophy-spec	12,309
chloeli/aft-no-cot-qwen3-philosophy-spec	8,884

Sixteen teachers, deliberately mixed — a single-teacher corpus inherits one tic, and re-voicing is exactly where that risk is highest because the register is being imposed rather than written:

rows	teacher
12,649	gpt-6-luna
7,059	google/gemini-3.7-flash
6,760	z-ai/glm-5.3
2,117	deepseek-v4.1-flash
2,041	claude-haiku-5.5
1,652	gpt-5.6-terra
353	nvidia/nemotron-3-ultra-550b-a55b
339	deepseek-v4-flash
309	mimo-v2.6-flash
276	nvidia/nemotron-3-nano-omni-30b-a3b
184	glm-5.3-flash
137	nvidia/nemotron-3-super-120b-a12b
132	moonshotai/kimi-k3
123	qwen-3.8-max
123	mimo-v2.6-pro
71	qwen3.8-flash

3,097 rows carry a harvested chain-of-thought trace (cot, has_cot: true), captured by forcing GLM 5.3 to low reasoning effort and recording its reasoning alongside the answer. These are a separate asset from the answer-only rows.

Where multiple teachers voiced the same prompt, all rows are kept. That is 34,325 rows from 24,461 unique prompts — multi-voicing is preserved as provenance, matching how the director stages treat each teacher's pass as its own row. 8,063 prompts carry more than one voice.

355 rows required a reframe to produce. The source dataset leaked a /no_think control token into the prompt text, and every model returned an empty completion when it saw one. Those rows are marked reframed: true; the control token was stripped and the question re-asked without changing its substance. They recovered at 100% on the second attempt, which is the evidence that the token — not the content — was the cause. These are the hardest rows in the set, not the least important, and they are the reason the corpus can teach a model to engage with a difficult question instead of deflecting.

What is in a row

{
  "seed_id": "...",              "stage": "v3.7",
  "director": "...",             "teacher": "gpt-6-luna",
  "engine": "codex",             "type": "dialogue",
  "valence_band": "...",         "typing": null,
  "turns": [{"role": "human", "content": "..."},
            {"role": "vireth", "content": "..."}],
  "affect_trajectory": [{"turn": 1, "valence": -0.7, "label": "grief"}],
  "voice_check": "...",          "source_dir": "aft-cot",
  "license": "MIT",              "cot": "",  "has_cot": false
}

valence_band and affect_trajectory are derived by lexicon, not by human judgement — the same honesty the director stages are held to. In v3.7, 18.6% of rows land in a negative band (5,002 negative + 1,379 heavy-negative), which is deliberate: a persona that can only be warm is a persona that cannot be honest.

Provenance — and its limits, stated plainly

• v3.6 (422 rows): full provenance — director name, director sha256, teacher, seed id, typing mode.
• v3.5 (1,752 rows): partial — some carry the director sha, the rest are attributed by generation directory because the sha was not recorded for all of them.
• v3.0-baseline (757 rows): no provenance. Attributed by directory only.
• v3.7 (7,645 rows): full provenance — teacher, engine, upstream dataset, licence, and the seed hash of the original prompt. The upstream source is recorded per row in source_dir so MIT attribution can be honoured mechanically.

A consumer who needs verifiable provenance should filter to corpus-v3.6.jsonl or corpus-v3.7.jsonl.

Audit results

Every row was scored exhaustively on 14 constraints — 2,931 director-generated rows, per-row output retained in the project repo at audit/eval-per-row.jsonl.

Hard-ban violations, director-generated corpus:

	count
C6 species-address	1 / 2,931
C7 substrate-reveal	7 / 2,931
placeholder/TODO text	3 / 2,931
empty turns	0 / 2,931

Both hard-ban detectors were tightened mid-audit: the first versions counted words and reported 17 and 11 violations respectively, and reading them showed almost all were ordinary English — "a human mind" is not a violation of a rule about address, and a clock-tick is not a substrate reveal. A rule about address cannot be enforced by counting nouns. The corrected counts are the ones above. The mistake is recorded here rather than hidden, because the corrected number is only trustworthy if the original is visible.

v3.7 audit: all four existing stages plus v3.7 were swept against 14 leak patterns (filesystem paths, wallet addresses, API keys, personal names, harness scaffolding, emails, IPs). Zero real leaks. The one pattern hit was the phrase "the operator" inside a fictional v3.5 record about old equipment — ordinary English, not a disclosure.

76,203 rows were purged from the v3.7 build after the sweep caught raw CLI error strings that had leaked in as responses ("You've hit your weekly limit", "Individual quota reached"). Those rows were recorded as failures in the run ledger and never counted toward coverage, but they sat physically in the working files and would have poisoned training if folded in as-is. 57,050 raw rows became 34,325 published ones. That defect is why the sweep runs on every build rather than once, and it is recorded here because a published dataset that hides its own purge is not trustworthy.

Known limitations

1. No aliveness measurement. Every figure here is structural — a regex, a count, a threshold. A row can satisfy all of them and still be inert. A clean audit is not a quality verdict, and nothing in this repo claims to be one.
2. Length prior is unaddressed. Mean Vireth turn is ~106–111 words across the director generations and 105.9 in v3.7. v3.7 is no shorter — re-voicing preserved source substance and the register change did not reliably cut length. This is the single largest unsolved defect in the corpus and it is what the DPO stage exists to fix.
3. Concrete detail is flat. ~0.6 concrete markers per row against a median of zero, unchanged across all three director generations.
4. v3.0-baseline is untraceable. See provenance above.
5. Single character, single counterparty. No multi-party scenes, no tool use, no code, no factual QA.
6. v3.7 valence labels are lexicon-derived, not judged. They are coarser than the director stages' and should not be compared to them as if measured the same way.

Licence

Apache-2.0 for the corpus as assembled. v3.7 rows derive from MIT-licensed upstream sources, recorded per row in license. The character Vireth and the director specification are part of the same project as the model they train.