README.md

metadata

pretty_name: Qever decision corpus and API distillation pilots
license: other
license_name: mixed-source-licenses
license_link: https://huggingface.co/datasets/qforge/qever/blob/main/SOURCES.md
task_categories:
  - text-classification
size_categories:
  - 1M<n<10M
tags:
  - decision-model
  - distillation
  - bdh
  - multilingual
  - source-derived
  - logprobs
configs:
  - config_name: sol-pilot
    data_files:
      - split: train
        path: sol-pilot/train-*.parquet
  - config_name: luna-pilot
    data_files:
      - split: train
        path: luna-pilot/train-*.parquet
  - config_name: source-candidates
    data_files:
      - split: train
        path: source-candidates/train/*.parquet
      - split: validation
        path: source-candidates/validation/*.parquet
      - split: calibration
        path: source-candidates/calibration/*.parquet
      - split: test
        path: source-candidates/test/*.parquet
  - config_name: sol-azure-100k
    data_files:
      - split: train
        path: sol-azure-100k/train-*.parquet
  - config_name: sol-azure-100k-jev
    data_files:
      - split: train
        path: sol-azure-100k-jev/train-*.parquet
  - config_name: sol-azure-300k-jev
    data_files:
      - split: train
        path: sol-azure-300k-jev/train-*.parquet
    default: true

Qever

Qever converts existing Hugging Face datasets into decisions of the form state + question + candidate options. It supports experiments with small decision models, including BDH students. This release includes approximately two million source-derived training candidates and an initial paired API annotation pilot.

The two-million-row corpus is not fully teacher-distilled. Luna and Sol were tested on the same 1,000 inputs. The published pilots retain the same 999 inputs after excluding one malformed source problem; these are not 1,998 unique problems. No newly trained model is included in this dataset release.

Configuration	Train	Validation	Calibration	Test	Teacher

| source-candidates | 1,999,998 | 25,951 | 25,254 | 24,624 | None; source gold where available | | luna-pilot | 999 | — | — | — | GPT-6 Luna | | sol-pilot | 999 | — | — | — | GPT-6 Sol |

The corpus covers evidence and entailment, paraphrase, rules, multilingual request routing, answerability and extraction, math, code behavior, chess tactics, safety judgments, scoring rubrics, response preferences, web elements, and tool selection. It combines human-authored and existing synthetic source material. See source attribution and transformations for all 19 pinned sources and their licenses.

A spot-check found an impossible math question that both teachers accepted. That source problem and all its derived decisions are excluded from the published files. See the quality exclusion. The teacher measurements below describe all 1,000 tested examples before this exclusion. Other source errors may remain.

API pilot measurements

The pilot deterministically samples 52–53 training examples per source and balances tasks within each source. Both teachers use reasoning.effort="none", temperature=1, and top_logprobs=20. The prompt requests one letter; billed output includes framing. Source labels are never sent to the teacher. These are annotation-agreement measurements on training examples, not held-out benchmark results or verified correctness.

Teacher	Mean input tokens	Mean billed output tokens	Complete candidate distributions	Source agreement	API cost for 1,000
gpt-6-luna	318.54	5.012	144/1,000	612/894 (68.5%)	$0.03529
gpt-6-sol	318.54	5.000	54/1,000	658/894 (73.6%)	$0.70564

Costs are calculated from returned API usage and the published Standard rates, including reported cache writes and hits. They are not invoice or account-balance readings. No Batch discount is assumed. Prices: Luna, Sol. See the detailed paired audit for per-source results and probability coverage.

Label semantics

• gold is the independently retained source answer index. It is null for teacher-only routing questions. Source-provided answers can be noisy, subjective, or based on synthetic solutions.
• teacher_prediction uses the highest-probability returned candidate code when available. teacher_emitted_prediction records the sampled visible answer separately. At temperature 1 these can differ. teacher_prediction_method records which rule was used.
• teacher_candidate_logprobs preserves returned logprobs in option order, with null for missing candidates. Missing values are never silently converted to zero.
• teacher_probabilities contains a normalized conditional distribution only when every candidate is present at the same verified initial output token position. Otherwise it is null. These probabilities are conditional on the candidate letter tokens and are not calibrated correctness estimates or AutoJev readout logits.
• teacher_observed_candidate_mass records the probability mass covered by returned candidate codes. teacher_unobserved_mass_upper_bound is the remaining mass calculated from the rounded API logprobs; it also includes non-candidate tokens. A small omitted mass does not establish that the predicted answer is correct.
• teacher_method distinguishes complete distributions from hard labels with incomplete logprobs. Preserve this distinction when training. Do not fill every missing probability with zero and call the result an exact teacher distribution.
• teacher_revision records the returned API model identifier, which was an alias in this run, not an immutable open-weight checkpoint. Server inference precision is undisclosed and left null.
• teacher_input_tokens_api and teacher_output_tokens_api are actual returned API usage. The older teacher_input_tokens column measures the AutoJev/Qwen prompt, while student_input_tokens measures the ModernBERT tokenizer. These tokenizers are not interchangeable.

Quality and evaluation boundaries

The source corpus is deduplicated by normalized decision content. Related questions are grouped by underlying passage, problem, conversation, game, rubric, or website. Exact states of at least 50 characters merge groups across sources. All groups are disjoint across the four published splits. Two short generic state hashes recur across splits and are reported in the audit.

Public JevBench states and the original BDH pilot evaluation states were used only for exclusion. The private exclusion text is not published. BBH task families and unsupported image encodings are excluded. These checks reduce known overlap; they do not establish that every upstream training mixture or API teacher is benchmark-contamination-free.

Most candidates fit the 1,024-token student and AutoJev prompt limits; longer rows are marked extended_4096. No over-limit prompt is silently truncated. Web decisions use text candidate projections, not screenshots. TACO examples use source input/output cases without executing code. Math source answers were not independently proved, and chess labels come from the source puzzle continuation. Subjective score disagreements require separate analysis from objective answer errors.

Teacher/source disagreements remain published. Do not replace trustworthy source labels with teacher predictions automatically. The pilots use no hidden reasoning; they do not test recurrent depth, reasoning distillation, zero-shot generalization, or student latency. The API pilot rows are training data and must not become a test set after training on them.

Load

from datasets import load_dataset

pilot = load_dataset("qforge/qever", "sol-pilot", split="train")
soft = pilot.filter(lambda r: r["teacher_probabilities"] is not None)
candidates = load_dataset("qforge/qever", "source-candidates", split="train", streaming=True)

Use gold for source-supervised examples where appropriate. Use complete teacher vectors for a soft-target loss; otherwise use an explicitly chosen hard-label or partial-logprob objective. Keep source agreement and annotation method available to the training pipeline.

Reproducibility

Selection audit before the final quality exclusion, file checksums, source registry, and pipeline code are included. Source revisions, source row identities, transformations, and licenses are also present per row. The API annotation date is 22 September 2026 UTC. API responses depend on provider model aliases and are not guaranteed to be exactly reproducible.

See SOURCES.md before redistribution or reuse. This collection retains mixed source licenses and underlying notices; it does not grant a new blanket license over the source content.

Azure GPT-6 Sol 100k development configuration

sol-azure-100k has 100,000 training-fold decisions selected from qforge/qever@e12c97a6d75bf74ba52dfb51747b43dc44d3266f. 99,049 were newly annotated through an Azure OpenAI Global Standard GPT-6 Sol deployment; 951 completed direct OpenAI pilot decisions were reused to avoid duplicate charges. The exact per-row provider, deployment, candidate logprobs, target method, source provenance/license, and quality status are retained. Exact distributions require all option letters at the same verified output token; approximate vectors require >=0.999 observed candidate mass and are explicitly separate from raw nullable logprobs. Source gold remains null where absent.

These are teacher labels, not independently verified gold or held-out test examples. review_required denotes source-answer uncertainty; the chess mechanical verifier does not establish tactical optimality. The separate validation/calibration/test splits stay in source-candidates; no evaluation text was relabeled. See sol-azure-100k/MANIFEST.json for checksums, source counts and conservatively priced Azure usage. Underlying source-specific licenses and attribution still apply.

Azure prompt-filter rejections: 307; ambiguous requests skipped without retry: 14. Their conservative reservations remain in the manifest's cumulative liability; the reported liability is an upper-bound estimate, not an invoice.

Training targets: 5,067 complete conditional option distributions; 75,108 bounded approximations (missing raw logprobs remain null, zero only in the derived training vector); 19,825 hard decisions. Independent source gold is unavailable for 10,527 rows.

Simple Jev-shaped Sol training view (default)

sol-azure-100k-jev is a three-column view of the same 100,000 Sol-labeled training decisions in sol-azure-100k, not another labeling run. Parse the input and output JSON strings with json.loads; using JSON strings preserves the option-to-probability maps as JSON objects in Parquet/Arrow.

• input is a TypeSafe Jev Choice request: state, model: "jev-latest", and questions.decision with type: "choice", instructions and ordered numeric-string criteria keys. It is a target API representation, not the request actually sent to Sol.
• output has Jev's model, answers.decision (type, choice, probabilities, confidence), and usage keys. It is not a Jev response: model: "gpt-6-sol" and token usage are from Sol, probabilities are Sol-derived training targets, and confidence is null because Jev never evaluated these inputs. Numeric Jev confidence cannot be manufactured. choice is the highest-probability option, which differed from Sol's sampled emitted letter for some rows; the emitted code is retained in metadata.
• metadata retains ID/group/split, source provenance and license, source gold (null when unavailable), quality flags, provider, target method, original decision primitive, emitted choice, and raw nullable Sol logprobs. exact probabilities are conditional on the offered options; approximate probabilities are bounded derived estimates with missing raw logprobs still null and zeros only in the derived vector; hard values are one-hot choices not measured probabilities or confidence.

Example (confidence: null is intentional):

{"input":{"state":"...","model":"jev-latest","questions":{"decision":{"type":"choice","instructions":"Which option?","criteria":{"0":"A","1":"B"}}}},"output":{"model":"gpt-6-sol","answers":{"decision":{"type":"choice","choice":"1","probabilities":{"0":0.0,"1":1.0},"confidence":null}},"usage":{"input_tokens":100,"output_tokens":5}}}

Load explicitly with load_dataset("qforge/qever", "sol-azure-100k-jev", split="train"); for each row use json.loads(row["input"]) and json.loads(row["output"]). The default subset is this simple view. The original detailed Parquet config remains available for full audit at sol-azure-100k@23af630dbf2b21fd5e3ce4ef228195a3bba5c0bc. See sol-azure-100k-jev/MANIFEST.json for frozen hashes. No BDH student or Jev model was run on this expansion, and Sol labels are not verified source gold. Source-specific licenses still apply.

300k typed GPT-6 Sol training decisions (default)

sol-azure-300k-jev has 300,000 train-only rows and exactly three columns: metadata, input (Jev request JSON string), output (Sol-derived Jev-style response JSON string). The original 100,000 Choice-formatted Sol rows from sol-azure-100k-jev are preserved byte-for-byte; 200,000 distinct new Sol labels add 70,000 Choice, 80,000 Noul, and 50,000 Score requests. Thus the combined input types are 170,000 Choice, 80,000 Noul and 50,000 Score.

The teacher is GPT-6 Sol, not Jev. Noul values are Sol-derived estimates of the yes option; Score values are probability-weighted, zero-based ordered rubric levels. Their underlying option logprobs may be incomplete: each row's metadata.target_method distinguishes exact conditional probabilities, bounded approximate probabilities, and one-hot hard labels. Missing raw candidate logprobs remain null in metadata. A hard value 0 or 1 is not a measured certainty. Jev confidence is null (unobserved), and Sol's actual model/usage are shown in the output. metadata retains pinned source license, group, source gold where independently available, quality flags, and the original rubric-option alignment. These are annotations, not verified gold.

No validation, calibration, test, or prior HF configuration was relabeled. See sol-azure-300k-jev/MANIFEST.json for hashes, methods, counts, and conservative cumulative Azure liability. Source-specific licenses and attribution continue to apply.