OusiaResearch/Vireth-DPO-Preference-Pairs

Vireth — DPO Preference Pairs

9,044 audited synthetic preference pairs for the Vireth persona project.

This is a separate, complementary dataset to Vireth — Persona Corpus. The persona corpus supplies broad affect-heavy conversational material; this release supplies contrastive pairs for the three defects that remained visible after SFT-style corpus work:

1. overlong answers,
2. mirroring the user's message length and shape,
3. abstract, category-level language in place of concrete particulars.

It is not a general instruction dataset, a benchmark, human preference data, or a claim that passing these structural checks makes a response good. It is a narrow synthetic DPO stage.

Contents

field	meaning
prompt	The reconstructed human turn/context supplied to the generator: a recorded situation plus one concrete detail.
chosen	Reply generated under the desired constraint.
rejected	Reply generated under the opposite failure instruction.
axis	LENGTH, ANTI-MIRROR, or CONCRETENESS.
teacher	Generation teacher/route. The chosen/rejected label is not assigned by teacher identity.
seed_id	Stable identifier for the recorded synthetic seed.
prompt_turn	Internal generation-slot trace ID.
chosen_words, rejected_words	Whitespace-token word counts for inspection, not tokenizer counts.
generated_utc	Generation timestamp when preserved by the source leg.

The main file is data/vireth-dpo-v1.jsonl (one JSON object per pair). data/DPO-STATS.json is the machine-readable build record.

Final composition

axis	pairs	share
ANTI-MIRROR	3,322	36.7%
CONCRETENESS	2,919	32.3%
LENGTH	2,803	31.0%
Total	9,044	100%

The set contains approximately 1,011,844 whitespace-token words across chosen and rejected responses (~1.32M heuristic tokens). It is 8.53 MB as JSONL.

Sixteen generation routes are represented. The largest contributors are Codex (1,999 pairs), Luna (1,495), Haiku (1,064), Claude Haiku 5.5 free (773), and DeepSeek V4.1 Flash free (705). Mixing teachers is deliberate: a single-teacher preference corpus would turn one model's habits into the definition of "preferred."

How pairs were made

Every pair begins with a synthetic seed containing a situation, a speaker texture, valence band, and concrete detail. The released prompt is reconstructed from the recorded situation and first concrete detail — the same context form passed to the pair generator. The seed record is retained locally as build provenance; this release contains only the trainable prompt and pair.

For each axis, the generator asks for two replies under opposed instructions:

• LENGTH: say what is true at the length it needs vs. an extended, overly interpretive response.
• ANTI-MIRROR: respond at the length warranted by the situation vs. copy the person's message shape and finish with a friendly assistant-style close.
• CONCRETENESS: name actual particulars — day, place, name, thing said — vs. describe generalized states and categories.

Acceptance is gate-verified by the relevant axis rather than by which route wrote it. For example, the concreteness gate requires specifics in the chosen side and an abstraction marker in the rejected side; the anti-mirror gate compares both replies to the reconstructed human turn. The release keeps the generator's structural labels, not a claim of human comparative judgement.

Audit and purge

The build started with 9,527 raw candidate pairs. 483 were excluded before release:

excluded	count	reason
CLI-error contamination	426	Provider quota/error strings appeared where dialogue should be; one raw pair had error text on both sides.
Chosen-side assistant signoff	39	The preferred side ended in an assistant-style closing.
Degenerate	16	One side had fewer than three words.
Personal-name leakage	2	Rows containing the operator's given name were removed.

The final file was then swept for filesystem paths, API-key forms, wallet addresses, personal names, harness text, agent names, email addresses, IPs, task identifiers, provider-error text, and leaked control tokens. The only remaining pattern hits are:

• 532 Mara matches: reviewed false positives; Mara is a fictional person in generated scenes, not an agent identity or operator reference.
• 100 signoffs on rejected: expected for the ANTI-MIRROR failure side; these model the unwanted assistant closer.
• 3 LENGTH pairs where simple word-count ordering is not separated: retained and marked in the stats because the axis's structural intent was still judged by its generation gate.

A clean leak audit is not a quality verdict. It only establishes that the released artifact does not contain the known classes of internal/operator contamination the build checked for.

Intended use

Use this as a DPO/ORPO-style preference component alongside a conversational SFT corpus, not as a standalone corpus. Preserve the axis field for ablations and diagnostics. Do not interpret the synthetic chosen/rejected labels as human preference data, safety ratings, or a universal standard of helpfulness.

The intended training order for the Vireth project is SFT → DPO → abliteration last. This dataset is the DPO stage only.

Limitations

• All labels are synthetic contrastive labels. They measure compliance with narrow structural instructions, not human taste or emotional truth.
• Prompts are compact reconstructed context, not full multi-turn conversations.
• The data is English-only and focused on a single persona/register.
• The corpus deliberately trains against generic assistance habits; that does not make every direct or brief response correct.
• No direct measurement here establishes downstream behavior or model quality.

License and provenance

Apache-2.0. The released prompts and pairs are synthetic. Teacher routes and seed IDs are retained for reproducibility and mixture analysis; no provider-error responses or identified operator-derived rows are released.