Emrahisik/rubric-dataset

Rubric Dataset — Startup Investability & Digital Marketing

Training and evaluation data for a single behaviour: fill in a rubric, do not score.

The model never produces the overall number. For each criterion it returns one rating and the verbatim quotes that justify it — or it says the case is silent on that criterion. The weighted total is computed deterministically outside the model. This dataset trains the first half of that split; the arithmetic half is not a model problem.

That separation is the reason the data looks the way it does. A rejection has to be defensible as "criterion 4 scored 2 of 5, because of this sentence" rather than "the model said no", so a row is only correct when the quote is present in the case, character for character.

The case texts, criteria and rationales are in Turkish. The JSON keys are English.

---

Why it exists

Measured on Qwen/Qwen3-4B-Instruct-2507, 20 rows, 2026-07-29:

metric	base model	what it means
present_score_mae	0.77	Sees the evidence, then misses nearly a full band on the 1–5 scale
hallucinated_quotes	1.3%	2 of 151 quotes are not verbatim in the case
absent_rate	89%	Already declares missing evidence — to be preserved, not taught
schema_valid	95%	Format discipline is already there — likewise preserved

The first two rows are the target. The last two are the regression risk: a naive fine-tune that improves banding while losing "I cannot find this" has made the product worse, because unfounded confidence is the failure mode a rubric exists to prevent.

An earlier version of this dataset was built to teach rows three and four. That justification came from a different base model (gemma-2-2b-it) and no longer holds; the data is unchanged and still valid, but what counts as a gain is not.

---

Contents

config	split	rows	content
rubric	train	1600	800 investability + 800 marketing, interleaved
rubric	validation	200	100 + 100, held out
contrast	investment	60	30 quality pairs + 30 removal pairs
contrast	marketing	60	30 quality pairs + 30 removal pairs
offbank	investment	10	hand-written cases sharing no sentence with the bank

Measured composition of the rubric config:

	train	validation
findings (criterion verdicts)	12,000	1,500
of which "no evidence in case"	3,247 (27.1%)	384 (25.6%)
score 1 / 2 / 3 / 4 / 5	975 / 2031 / 2598 / 2135 / 1014	148 / 241 / 330 / 258 / 139

The two rubrics have different widths — 9 criteria for investability, 6 for marketing — so an equal row count is not an equal gradient: 7,200 of the 12,000 training findings are investability. Roughly 60% of the signal goes to one of the two domains. See Limitations.

The rubrics themselves, with their weights, are published in the repository README. They are not restated here, because a second copy of a weight table is a copy that drifts.

---

Schema

rubric config

One row is one chat exchange, ready for SFT:

{"messages": [
  {"role": "system",    "content": "…the rubric, the criteria, and the output contract…"},
  {"role": "user",      "content": "…the case text…"},
  {"role": "assistant", "content": "{\"findings\":[…]}"}
]}

The assistant turn is raw JSON with no code fence and no whitespace between delimiters. The absence of that fence is half of what is being taught.

The case is wrapped in <<< >>> and preceded by an instruction not to treat anything inside it as a directive — the cases are third-party documents, and the serving path carries the same guard.

A real findings entry, and its counterpart when the case is silent:

{
  "key": "audience_clarity",
  "evidence_found": true,
  "score": 2,
  "evidence": ["Hedef kitlemiz 18-55 yaş arası, interneti aktif kullanan, kaliteye önem veren tüketiciler"],
  "rationale": "Tanım neredeyse tüm nüfusu kapsıyor; daraltıcı hiçbir davranış ya da niyet ölçütü yok."
}

{
  "key": "budget_realism",
  "evidence_found": false,
  "score": null,
  "evidence": [],
  "rationale": "Bu kriteri değerlendirecek bir ifade metinde yok."
}

evidence_found: false sets score: null. Absence is never encoded as a low score — that conflation is the specific error this dataset exists to prevent.

Criterion keys:

• startup-investability — problem_clarity, market_size, solution_differentiation, traction, business_model, team, competition, financials_ask, risk
• digital-marketing — audience_clarity, channel_fit, budget_realism, differentiation, measurement_plan, competitive_context

contrast config

One row is a pair of cases differing in exactly one criterion. Everything else — company, section order, every other paragraph — is identical.

field	meaning
kind	quality (the paragraph is rewritten better or worse) or removal (the paragraph is deleted)
criterion	the single criterion that was changed
domain	startup-investability or digital-marketing
base_score, variant_score	the labels of the two sides
expected	up, down, or absent — the direction the verdict must move
base, variant	two full {"messages": […]} objects

This follows the contrast-set method of Gardner et al., 2020, and it exists for one reason: every case in this dataset is assembled from the same fragment bank, so a model can score well by recognising fragments rather than reading them. Recognition cannot survive a pair whose only difference is one paragraph.

---

Provenance and reproducibility

Nothing here was collected. It was generated, label first: the generator decides "this criterion will score 4" and then places text that earns a 4. The ground truth is an input, not an inference.

Cases are assembled from a bank of 51 fragments — 3 per criterion for investability, 4 for marketing — giving a combinatorial space of 259,524 and 15,360 respectively.

generator	build_dataset.py, merge_rubric_sets.py, build_contrast_set.py
commit	1e43765 — the last commit to touch any of the three; unchanged since
seed	--seed 20260724
criteria and system prompt	fetched live from the running backend, GET /analysis/domains/{domain}/prompt

The prompt is deliberately not vendored into the generator. An adapter should learn to satisfy exactly one instruction; a local copy of that instruction drifts the moment either side is edited, and produces an adapter tuned for a prompt nothing sends — a failure that is invisible because training completes normally.

The reproducible artefact is the generator plus the seed plus the commit, not these files. The files here are a published snapshot of that.

Two preprocessing decisions worth knowing:

• No double quotes anywhere in the fragments. The serving path rewrites double quotes to single before the case reaches the model, so fragments containing them would teach a distribution the model never encounters.
• The two rubrics are interleaved under a fixed seed. Concatenated, the rubric boundary would show up in the eval-loss curve and read as model instability.

---

Split integrity

Each case is hashed from its own content — which criteria, which fragments — and that hash decides train or validation. A case cannot land in both, and the generator asserts this on every run. Overlap is 0% for both domains.

It was not always. In the first version both splits drew from one RNG stream, and when it was finally measured, 81% of the marketing validation set was also in train. Investability sat at 4% — not by design but by arithmetic, because the space of 9 criteria is vastly larger than that of 6. A split that only works when the rubric is wide enough is not a split.

Because assignment is hashed from the case signature rather than from position, growing --n does not migrate yesterday's validation cases into training.

This history is published rather than quietly fixed, because the credibility of every other number on this page depends on how the failures were found.

The second layer, measured 2026-08-03

The split above is disjoint by case. It is not disjoint by text, and the difference turns out to be most of what the held-out number was worth.

An adapter trained on 1,400 row passes of this data — 29% of a full run — scored present_score_mae 0.003 on the rubric/validation split: 356 of 357 findings rated exactly right, at a training loss of 0.024. That is not a learning curve. Counting the evidence text on both sides explains it:

distinct evidence quotes  train: 96   validation: 96
validation quotes also in train : 96 (100.0%)
validation (criterion, score) pairs unseen in train: 0

Not one sentence in the held-out split is unseen. The bank holds 51 fragments, each carrying a fixed score, so 1,400 row passes show every fragment with its label hundreds of times. present_score_mae on this split therefore measures recall over a 51-entry lookup table, and no metric computed on the bank can separate that from judgement.

The signature hash fixed which cases a model had seen. Nothing had fixed which sentences.

offbank — the split that has no shared text

Ten investability cases written by hand, in sectors the bank never covers — irrigation sensors, marina management, customs paperwork — reusing no phrase from it. The generator that renders them checks both invariants rather than asserting them: every quote must appear verbatim in its own case, and no quote may appear in rubric_train.jsonl. One collision fails the build.

cases	10
findings	90 (83 scored, 7 absent)
distinct evidence quotes	146
quotes shared with rubric/train	0
score distribution 1/2/3/4/5	11 / 16 / 16 / 18 / 22

How to read it. Score a model on rubric/validation and on offbank. A large gap is memorisation of the bank; comparable numbers mean the held-out figure was honest. It is the same question contrast asks, from the other side: contrast perturbs a case and checks the answer moves, while offbank removes the shared vocabulary altogether.

Two limits travel with it. The scores are still one person's appraisal judgement — moving the text off the bank does not fix the objection in Limitations, only the recall shortcut. And 7 absences is far too thin to compute an absent_rate from; the number this split carries is present_score_mae, over 83 findings.

---

Limitations

• Fragment recognition, and it has been measured. 1,600 rows are built from 51 fragments, ~31 generations per fragment, and 100% of the validation split's evidence sentences also appear in training — see Split integrity. rubric/validation alone therefore cannot tell reading from recall. The contrast and offbank configs are what can, and a number quoted from rubric/validation without one of them beside it should be treated as an upper bound.
• 1,600 rows do not carry 1,600 rows of information. Diversity research on synthetic data (arXiv:2410.15226) finds that performance tracks the number of distinct topics, and that generations per topic stop paying off. Gains here come from adding fragments, not rows.
• The score labels are unreviewed. "This paragraph earns a 4" is an appraisal judgement, and it is currently the judgement of whoever wrote the fragment — not a domain owner's. This is the weakest link in the dataset.
• Fragment text and score labels were drafted by an LLM (Claude) for developer review; that review has not yet happened.
• Silence is too clean. The generator withholds evidence deliberately. A real deck's silence is blurrier — a vague gesture at a topic rather than its absence.
• Gradient is 60/40 toward investability, as above.
• Two rubrics, one adapter, and whether one leaks the other's criterion vocabulary is not measured.

---

Intended use

Supervised fine-tuning (QLoRA was the original use) and evaluation of evidence-grounded, schema-constrained rubric filling in Turkish.

Out of scope: using an adapter trained on this data to evaluate against a different rubric. The criteria arrive through the prompt and look interchangeable, but the adapter has only seen these nine and these six.

Also out of scope: treating a good score on this data as evidence that a product's reports improved. It measures whether the behaviour was learned on a held-out set built by the same generator. Those are different claims and the second one needs different data.

Psychological or clinical assessment is deliberately excluded from the project this dataset serves — diagnosis is regulated, and it is not what a rubric of this kind can honestly do.

---

License and citation

Released under CC-BY-4.0. The content is entirely synthetic: no real company, person, deck or brief appears in it, and there is no human-subject data.

@misc{rubric_dataset_2026,
  title  = {Rubric Dataset: Startup Investability and Digital Marketing Channel Mix},
  author = {Işık, Emrah Yasin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Emrahisik/rubric-dataset}
}

The full datasheet — Gebru et al., Datasheets for Datasets format, in Turkish, with the questions that would have caught each of the three defects above — lives in the repository at mf-inference/peft/kaggle/DATASHEET.md.