README.md

metadata

license: apache-2.0
task_categories:
  - text-classification
tags:
  - llm-routing
  - semantic-routing
  - semantic-features
  - probabilistic-modeling
  - interpretable-ai
  - selmroute
  - jev
  - laya
configs:
  - config_name: jev
    default: true
    data_files:
      - split: full
        path: data/performance/features_jev.csv
  - config_name: laya
    data_files:
      - split: full
        path: data/performance/features_laya.csv
  - config_name: jev-compact12
    data_files:
      - split: full
        path: data/performance/features_jev_compact12.csv
  - config_name: jev-cost
    data_files:
      - split: full
        path: data/cost/features_jev.csv

SeLMRoute Semantic Features

Frozen numerical semantic features for SeLMRoute: Probabilistic Semantic Evidence for Large Language Model Routing, by Vasilis Perifanis, Nikolaos Pavlidis and Symeon Symeonidis.

These files are copied unchanged from the source repository at commit ec42fdbad3d30320679f0a25a7f7df335d379f1d. No semantic extraction was rerun and no CSV values or column names were changed.

Configurations

Configuration	Original file	Rows	Columns
jev	data/performance/features_jev.csv	11,481	112
laya	data/performance/features_laya.csv	11,481	112
jev-compact12	data/performance/features_jev_compact12.csv	11,481	73
jev-cost	data/cost/features_jev.csv	12,446	112

Each configuration has one split named full, containing its complete frozen pool. This is a packaging label, not a training/test assignment. Reproduce grouped evaluation with the source project's splitting code; do not treat these complete pools as held-out test sets for the full-data deployment checkpoints.

The performance JEV and Laya files cover the same 11,481 sample IDs. The cost file is a separate 12,446-sample pool. The compact12 file contains its own frozen reduced-probe extraction and cannot replace the 40-input feature file for either released deployment checkpoint.

Load with Hugging Face Datasets

from datasets import load_dataset

repo = "Indigma-Innovations/SeLMRoute-Semantic-Features"
jev = load_dataset(repo, "jev", split="full")
laya = load_dataset(repo, "laya", split="full")
compact = load_dataset(repo, "jev-compact12", split="full")
cost = load_dataset(repo, "jev-cost", split="full")

load_dataset(repo) selects the default jev configuration. Numerical CSV data can also be loaded directly with pandas. No TypeSafe API key or Laya model is required to read these frozen features.

Schema and router input

Each row is keyed by a deterministic sample_id. column_schema.json lists every column in source order for each configuration. schemas/jev_router_metadata.json and schemas/laya_router_metadata.json preserve the exact ordered 40 inputs and 20 outputs of the corresponding deployment routers.

112 CSV columns do not mean 112 router inputs. The full files contain an identifier, deterministic text statistics, semantic probability distributions, auxiliary summaries, a diagnostic task-family distribution and extraction metadata. Only the 40 names in the matching router metadata form its ProbabilityMass input.

Column pattern	Meaning
sample_id	Join key to the matching benchmark release; not a router feature.
det_*	Deterministic text statistics and indicators; excluded from these deployment checkpoints.
jev_<probe>__noul / laya_<probe>__noul	Binary-probe probability retained as one input per binary probe.
<backend>_<probe>__p__0 through __p__3	Four-level probability distribution, in criterion order.
<backend>_<probe>__entropy	Normalized entropy derived from the distribution.
<backend>_<probe>__confidence	Confidence supplied by the semantic backend.
<backend>_<probe>__score	Score supplied by the semantic backend.
<backend>_<probe>__expected	Expected level calculated from the distribution.
<backend>_<probe>__spread	Standard deviation calculated from the distribution.
<backend>_task_family__*	Auxiliary task-family evidence, excluded from the 40-input representation.
meta_<backend>_input_tokens	Recorded semantic-extraction input-token metadata.
meta_<backend>_latency_ms	Recorded semantic-extraction latency metadata; zero values must not be interpreted as a zero-cost live service.

The compact12 configuration has fewer probe columns. Frozen values retain source precision and rounding; do not regenerate or normalize them when reproducing the study.

Semantic probes

The full representation consists of eight binary probabilities plus eight distributions over levels 0–3: 8 + 8 × 4 = 40. The exact probe questions and ordered level criteria are included in configs/probes_manual_v1.yaml. That source file also defines an auxiliary task_family diagnostic, excluded from the 40 inputs.

Probe	Type	Meaning
math_reasoning	Binary probability	Does correctly solving query require mathematical calculation, symbolic manipulation, or quantitative reasoning beyond copying an explicitly stated value?
code_reasoning	Binary probability	Does correctly solving query require writing, modifying, debugging, or reasoning about executable computer code?
formal_logic	Binary probability	Does solving query materially require formal logical, combinatorial, rule-based, or constraint-satisfaction reasoning?
factual_recall	Binary probability	Can query be answered primarily through factual knowledge or retrieval, with little derivation or multi-step reasoning?
social_affective	Binary probability	Does correctness materially depend on interpreting emotion, intention, social context, interpersonal meaning, or conversational affect?
tool_interaction	Binary probability	Does completing query require interaction with an external tool, software environment, API, or simulated environment rather than only producing an answer?
external_knowledge	Binary probability	Does solving query require factual information that is not explicitly supplied in the query and cannot be derived from the supplied information alone?
current_information	Binary probability	Would correctness materially depend on recent, changing, or time-sensitive information?
domain_specialization	Four-level distribution	How specialized is the knowledge required to solve query correctly?
reasoning_depth	Four-level distribution	How much sequential reasoning is required to reach a correct answer to query?
constraint_density	Four-level distribution	How many independent requirements or constraints must simultaneously be satisfied by a correct answer to query?
context_integration	Four-level distribution	How much information from different parts of query must be integrated to solve it correctly?
decomposition_need	Four-level distribution	To what extent does solving query require decomposing it into distinct intermediate subproblems?
ambiguity	Four-level distribution	How underspecified or semantically ambiguous is query?
exactness	Four-level distribution	How sensitive is correctness to exact details, precise constraints, exact values, or exact output behavior in query?
answer_openness	Four-level distribution	How broad is the set of answers that could reasonably count as correct for query?

Use with the released routers

• SeLMRoute-JEV: use the jev configuration.
• SeLMRoute-Laya: use the laya configuration.

The model cards provide runnable inference examples. Read feature_names from the matching model's metadata.json rather than reconstructing feature order or passing all CSV columns. These checkpoints predict scores for the 20-model performance pool; the jev-cost configuration supports a different experiment in the source code.

Benchmark provenance and targets

The features are derived from LLMRouterBench release:

• Dataset: NPULH/LLMRouterBench
• Revision: 0e5af1b84bf73437a01a1849c0f1d2468baa93fc
• Bundle: bench-release.tar.gz
• SHA-256: b79f8cde1a6f029c2efa663a3a3b6f7748defb22341fe59f328cebef6648c8f1

This release contains numerical features and IDs, not raw questions, answers, model completions or training targets. Obtain the frozen benchmark bundle separately to reconstruct samples.csv and outcomes.csv and join on sample_id. Run the following from the source repository:

uv run selmroute prepare-data --bundle /path/to/bench-release.tar.gz

The performance pool spans 15 datasets and 20 models; the separate performance-cost pool spans 10 datasets and 13 models. See THIRD_PARTY_DATA.md for the constituent benchmarks and distribution boundary. Full paper reproduction also requires the other artifacts described in the source README, including the separately hosted embedding baseline; these four feature files alone are not the complete reproduction bundle.

Limitations and license

Semantic features reflect the frozen extractors and benchmark distribution; they are not ground-truth annotations of task requirements. JEV and Laya values are not interchangeable. Dataset shifts, extractor revisions and changing candidate models require separate validation. Group related or repeated query inputs according to the source evaluation protocol to avoid leakage.

Apache-2.0 applies to original SeLMRoute material. It does not replace the independent terms of LLMRouterBench, its underlying benchmarks or third-party model providers. See LICENSE, NOTICE and THIRD_PARTY_DATA.md.

Citation

@article{perifanis2026selmroute,
  title = {SeLMRoute: Probabilistic Semantic Evidence for Large Language Model Routing},
  author = {Perifanis, Vasilis and Pavlidis, Nikolaos and Symeonidis, Symeon},
  journal = {arXiv preprint arXiv:2609.34736},
  year = {2026},
  url = {https://arxiv.org/abs/2609.34736}
}