frangelbarrera/ling-3-flash-evaluation

Ling-3.0-flash Evaluation — Sanitized Records

This dataset is a sanitized, structured release of the independent evaluation of inclusionai/ling-3.0-flash:free conducted through the OpenRouter API. It contains one metadata record per evaluation call and is intended for transparent analysis of model behavior across reasoning, coding, security, tool calling, long context, multi-turn stability, multilingual behavior, hallucination, format following, and edge cases.

The canonical research repository is frangelbarrera/Ling-3-flash-evaluation. The interactive presentation is maintained as a separate static Hugging Face Space linked from this dataset.

Dataset summary

The release contains 845 records from 12 test phases. The records preserve evaluation metadata and outcome signals while intentionally excluding raw prompts, system messages, model responses, tool payloads, model reasoning traces, observations, and error text.

Property	Value
Records	845
Evaluation target	inclusionai/ling-3.0-flash:free
API route	OpenRouter
Evaluation phases	12
Standard-schema records	692
v6 message-schema records	101
v6 message/tool records	52
Completed responses	680
API errors	63
Empty responses caused by length	52
Truncated responses	25
Other empty responses	25
Output format	JSONL and CSV

Intended use

This dataset is intended for research and education involving LLM evaluation, reproducibility, model reliability, safety behavior, tool-calling validation, long-context experiments, and cybersecurity-oriented capability analysis. It is suitable for building aggregate dashboards, comparing evaluation phases, studying failure classes, and demonstrating dataset documentation practices.

It should not be treated as an official benchmark, a leaderboard result, a safety certification, or a complete measurement of Ling-3.0-flash capability. Several phases are exploratory or small-sample behavior studies, and rate limiting affected some API calls.

Data fields

Each row contains stable identifiers, phase and test metadata, model/provider identifiers, parameter summaries, status and finish information, token counts, latency, and boolean indicators such as whether visible content, reasoning, or tool calls were present. The dataset does not contain the original textual prompt, model output, tool arguments, reasoning text, or error message.

The schema_variant field identifies whether a record originated from the standard JSONL schema, the v6 message schema, or the v6 message/tool schema. The result_class field normalizes completion, truncation, empty-output, and API-error outcomes without reproducing the original response text.

Sanitization policy

The following fields were excluded from this public release: user prompts, system prompts, message histories, visible model responses, reasoning traces, tool definitions, tool calls, observations, and raw error text. The release also excludes raw URLs, email-like strings, and other unstructured content because the source logs contain cybersecurity prompts, adversarial examples, and model-generated text that require separate review before redistribution.

A deterministic redaction report and field schema are included with the release package. Raw logs remain outside this release and are not required for the dataset viewer.

Reproducibility

The source repository documents the API configuration, evaluation sessions, parameters, known rate-limit effects, schema variations, and phase-level methodology. The source JSONL file hashes are recorded in the local redaction report so that the sanitized records can be audited against the exact source snapshot before publication.

The source project used seed=42 and temperature=0 for the v6 JSONL tests. Earlier sessions used different settings, which are preserved in the parameter fields and source documentation. No local model weights were used.

Limitations

The dataset reflects one model identifier accessed through a provider proxy and may include provider-specific behavior. Latency includes proxy overhead. The :free route may not represent every paid or future model deployment. Small samples and exploratory prompts produce wide uncertainty intervals. Evaluation timestamps, prompts, and raw responses are intentionally not included in this sanitized release, so this dataset is not sufficient by itself to reproduce every semantic judgment.

Licensing and terms

The source repository is licensed under MIT. This sanitized dataset is a derived research artifact containing metadata about API evaluations and is marked other pending confirmation of the applicable terms for redistribution of derived evaluation metadata and model/provider identifiers. The source repository's license should not be interpreted as automatically granting rights to redistribute all upstream prompts, provider responses, or third-party content. Publication should proceed only after the maintainer confirms the final license and applicable API/provider terms.

Citation

If this dataset is published, cite the source repository and the Ling-3.0-flash evaluation report:

@software{barrera_ling3_evaluation,
  author = {Barrera, Frangel},
  title = {Ling-3.0-flash: Independent Security and Capability Evaluation},
  year = {2026},
  url = {https://github.com/frangelbarrera/Ling-3-flash-evaluation}
}