SZLHOLDINGS/WILLAY

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SZLHOLDINGS/WILLAY is a 0.5 billion parameter research fine-tune based on Qwen/Qwen2.5-0.5B-Instruct, designed for SZL self-description with a 32768 token context length. Its training objective is to produce SZL-focused descriptions, but it does not establish reliable doctrine adherence, refusal behavior, or factual accuracy. Historical evaluations show a failure to pass held-out cases and refusal regression, with deployment remaining blocked due to unestablished artifact lineage and lack of production authorization.

Loading preview...

WILLAY Overview

WILLAY is a 0.5 billion parameter research fine-tune developed by SZLHOLDINGS, specifically aimed at generating SZL-focused self-descriptions. It is based on the Qwen/Qwen2.5-0.5B-Instruct model and supports a 32768 token context length. The model's primary purpose is research into self-description, but it is explicitly noted that it does not guarantee reliable doctrine adherence, refusal behavior, or factual accuracy.

Key Characteristics and Limitations

  • Research Fine-tune: Developed for specific research into SZL self-description, not general-purpose use.
  • Unreliable Output: The model's output should not be assumed as evidence; it does not establish reliable doctrine adherence or factual accuracy.
  • Failed Historical Evaluation: Past evaluations (September 2026) recorded a failure to pass held-out cases (0 of 5 passed) and refusal regression, leading to status=FAIL and publicationAuthorization=false.
  • Deployment Blocked: The model's deployment remains blocked due to unestablished artifact lineage, consent, privacy, training suitability, and lack of production authorization (production_ready=false).
  • Adapter-Based: The model includes both a root adapter for Qwen/Qwen2.5-0.5B-Instruct and a separate adapter for unsloth/qwen2.5-0.5b-instruct-unsloth-bnb-4bit.
  • No Verified Reproducibility: Gaps in merge lineage and unsigned receipts limit exact reconstruction and producer authentication.

Good for

  • Academic Research: Exploring the challenges of self-description in language models.
  • Understanding Model Limitations: Studying cases where models fail to meet reliability and factual accuracy standards.
  • Investigating Model Provenance: Analyzing issues related to artifact lineage, reproducibility, and deployment authorization in AI models.

Finetunes
1 models