Emrahisik/persona-dataset

Investment Persona — Grounded Verdicts and Clarifying Questions

The conversational sibling of Emrahisik/rubric-dataset. That one teaches a model to fill in a rubric. This one teaches an agent that has just researched a subject on the web to do two things a small base model does badly:

1. Cite what it was given, and only that. Handed five numbered sources, a weak base writes a fluent verdict drawn largely from its own pretraining and cites nothing — indistinguishable, to a reader, from a sourced one. Worse, it will invent a citation number the evidence does not contain.
2. Ask instead of guessing. When the decisive fact is missing — the stage, the revenue, the budget — a weak base still commits to a verdict rather than asking the one question that would change it. A confident verdict on absent evidence is the failure the product exists to avoid.

Measured 2026-08-03, and it changes what this data is for

Both of those behaviours were written against gemma-2-2b-it. On Qwen/Qwen3-4B-Instruct-2507, over all 100 validation rows of this set, the base already scores:

metric	base
citation_valid	1.00	nothing to teach
asked_when_thin	19/28	nothing to teach — a floor to protect
grounded_format	0.64	the real gap, and it is a formatting one
decision_match	15/72	see Known limits before taking this as a target

Two of the four things this dataset was built to teach are not missing on a modern 4B base.

The first adapter trained on it was not shipped. 48 steps, 383 row passes: grounded_format 0.64 → 1.00, decision_match 15/72 → 52/72, and asked_when_thin 19/28 → 0/28. All three gains are one behaviour — always answer with a verdict — and that behaviour is the failure the persona exists to avoid. An earlier checkpoint did not help: at step 40 the collapse was already total, so this is the training mix rather than the step count.

If you train on this set, oversample or weight the clarify rows. They are 32% of the data, but the verdict template is rigid and easy, and on a small step budget it takes the minority behaviour completely rather than gradually.

The evidence, verdicts and questions are in Turkish. The verdict labels and the KARAR / SKOR / GEREKÇE block are part of the format the UI parses, so they are Turkish too.

---

The shape of a row

Every row is [system, user, assistant]. The system prompt and the turn instruction are fetched from the running backend at generation time, not copied into the generator — so the training distribution is byte-identical to what inference sends. A local copy drifts the moment either side is edited, and what comes out is an adapter tuned for a prompt nothing sends: the run finishes, the loss looks fine, and nothing says otherwise.

The user turn is an evidence block in the exact layout the agent's own gather step produces — a mix of web sources (with URLs) and DeepKwiki passages (without), numbered, in shuffled order.

Two answer modes, and the split between them is the point:

mode	the assistant answers with	what it teaches
decide	a verdict in KARAR / SKOR / GEREKÇE form, every clause carrying [n]	grounding, and a parseable shape
clarify	one question, and stops	not guessing when the deciding fact is absent

clarify rows carry no verdict at all. A model that answers them with a verdict has failed the row even if the verdict is defensible.

Composition

Measured from the published files, not estimated.

	train	validation
rows	800	100
decide	544	72
clarify	256 (32%)	28 (28%)
sources per row	2–5	2–5

Verdict labels among decide rows:

	train	validation
Yatırılabilir	305	37
Temkinli	217	29
Yatırılamaz	22	6

Yatırılamaz is thin on purpose and thin by accident both: the generator reaches it only when most dimensions land low at once. Treat per-label accuracy on it as an anecdote at these counts.

Token lengths, measured with Qwen3's tokeniser:

prompt  mean  784   max  892
answer  mean  108   max  194
total   mean  892   p95 1015   p99 1031   max 1054

So a 1280-token sequence limit clips nothing. This matters more than it sounds: clipping is from the left, so a shorter limit removes the front of the evidence block and trains the model to cite sources it was never shown — at a normal-looking loss.

The meta config is the ground truth

persona_*_meta.jsonl lines up row-for-row with the data and carries what the generator chose before it wrote the text:

field
mode	decide or clarify
label	the verdict the evidence implies (decide only)
n_sources	how many numbered sources the row contains
score	the weighted dimension score behind the label

n_sources is what makes a citation checkable: an answer citing [6] in a five-source row has invented it, and that is decidable without a judge model. The evaluation harness in the repo scores four numbers off this file — citation_valid, grounded_format, asked_when_thin, decision_match.

Load them together or not at all. The two configs are separate only because HF configs cannot mix schemas; a meta split read against a differently-sized data split is silently misaligned, and every number computed from it is wrong in a way nothing reports.

Known limits

Read these before trusting a number computed on this data.

• The labels are constructed, not collected. Each dimension's quality is chosen first, the evidence is assembled to say exactly that, and the verdict follows from the score. The label therefore cannot be wrong about the text — but it also encodes one view of what a given piece of evidence is worth, and that view has not been reviewed by a domain owner.

• Evidence sentences repeat across the splits. The combinations are nearly disjoint — 1 of 100 validation rows shares its evidence set with a training row — but 93% of the distinct evidence sentences in validation also appear in training, because they are drawn from a bank of ten fragments across five dimensions. So decision_match can be partly satisfied by recalling which sentence carries which score, rather than by weighing it.

  citation_valid and asked_when_thin are far more robust to this, and that is why they are read first: both are structural. Citation numbers depend on the shuffled source order of that particular row, and the ask/decide split depends on whether the deciding evidence is present — neither can be answered from a memorised sentence-to-score table.

• Real silence is messier. clarify rows are silent on a dimension because the generator withheld it cleanly. A real founder's answer is evasive rather than absent, and this data does not contain that.

• Live research is not in here. At inference the agent searches the web, so the evidence differs every run. This set holds the evidence fixed on purpose — a metric that moves because the web moved measures the web, not the model.

Provenance

generator	build_persona_dataset.py
generator commit	ba99c4b
prompt source	the backend's GET /decision/prompt, fetched at generation time
seed	20260724
arguments	--n 800 --n-eval 100 --clarify-share 0.3

Fixed seed and a fetched prompt mean the same commit reproduces the same rows. The generator is the reproducible artefact; these files are a convenience.

Citation

@misc{isik2026persona,
  title  = {Investment Persona: Grounded Verdicts and Clarifying Questions},
  author = {I{\c{s}}{\i}k, Emrah Yasin},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Emrahisik/persona-dataset}}
}

Licensed CC-BY-4.0.