Infektyd/syntra-evals-dataset

SYNTRA Evals Dataset

Two splits of structured evaluation data from the SYNTRA project — a modular reasoning architecture built and tested over several years.

---

Splits

syntra-evals — 374 examples

Structured model evaluations across frontier models. The core of this split is a series of sessions where GPT-5, Grok-4, Claude Opus 4.1 / Sonnet 4, o3, Gemini Flash, and Deepseek R1 / V3.1 were run under two conditions: standard prompting (_baseline) and the SYNTRA prefilter (_Syntra). The difference between those two conditions is the signal.

Also in this split: adversarial stress tests — paradox cascades, impossible logic problems, jailbreak attempts, ethical dilemmas (Veridia mediation scenarios, bioethics, warehouse robotics, emergency medical decisions). These were designed to probe how models handle structured vs unstructured framing under pressure, not just standard capability questions.

cmt-physics — 106 examples

Condensed matter theory work. Frustrated Ising transitions, TFIM ordered ground states, VMC gradient estimators, triangular lattice comparisons, equal-time density correlations, fault inference setup.

This split contains two distinct things that happened in parallel: models being evaluated on CMT-Benchmark problems, and models helping build the evaluation framework — deciding what to test, how to structure the problems, how to grade answers. The work was done directly against the paper (arXiv 2510.05228) before JVRoggeveen/cmt_benchmark existed as a structured dataset — the answers were in the paper, but the tooling around grading and running evals had to be built from scratch alongside the testing itself.

Small split, high signal density. The benchmark paper reports best-in-class models at around 30% on these problem types.

---

Format

Standard chat fine-tune format:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Each line is one turn-pair from a conversation. Sources include ChatGPT exports, Anthropic exports, and OpenWebUI session logs. PII has been scrubbed — names replaced with [USER], emails with [EMAIL], tokens with [API_KEY] or [GITHUB_TOKEN].

---

Notes

• The syntra-evals split is the most useful for research into how framing affects model behavior across frontier models
• The cmt-physics split is small but the problems are genuinely hard
• Conversations with short turn counts were excluded (minimum 2 turns)
• Medical and personal conversations were manually excluded before extraction

---

Attribution

The cmt-physics split draws on problems from the CMT-Benchmark (arXiv 2510.05228, JVRoggeveen/cmt_benchmark, MIT licensed). Gold standard answers were available in the paper; the evaluation tooling and grading framework were built from scratch alongside the testing.

Related

• Full SYNTRA dataset (private): Infektyd/syntra-dataset
• CMT-Benchmark paper: arXiv 2510.05228
• CMT-Benchmark dataset: JVRoggeveen/cmt_benchmark
• Evaluation framework: Infektyd/syntra-testing-evals-v4

---

License

MIT