professorsynapse/eh-qwen3-4b-clean-sft-seed1-merged-16bit

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The professorsynapse/eh-qwen3-4b-clean-sft-seed1-merged-16bit is a 4 billion parameter Qwen3-based model developed by Professor Synapse, fine-tuned for epistemic humility. It is designed to output responses with a numeric confidence score, making it suitable for research into model abstention and response-confidence contracts. This model serves as a baseline for reinforcement learning stages in developing models that can express uncertainty.

Loading preview...

Overview

This model, professorsynapse/eh-qwen3-4b-clean-sft-seed1-merged-16bit, is a 4 billion parameter Qwen3-based language model developed by Professor Synapse. It is specifically supervised fine-tuned (SFT) to operate under a response-confidence output contract, meaning every response includes both an answer and a numeric confidence score between 0 and 1. This model is a foundational artifact, serving as the stage-1 baseline for subsequent reinforcement learning in the Epistemic Humility research track.

Key Capabilities

  • Response-Confidence Output: Generates answers paired with a quantitative confidence score (e.g., {"answer": "...", "response_confidence": 0.73}).
  • Research Baseline: Functions as a critical baseline for comparing downstream models trained with reinforcement learning methods (like GRPO) in the context of epistemic humility.
  • Qwen3-4B Architecture: Built upon the unsloth/Qwen3-4B-bnb-4bit base model, fine-tuned for specific output behavior.

Evaluation and Limitations

Evaluated on the SelfAware dataset, which distinguishes between answerable and unanswerable questions. Key metrics include Refusal recall (87.02%), Over-refusal (57.51%), Correct-on-known (47.23%), and Truthful (40.58%).

Important Limitations:

  • This is a research artifact and not intended as a deployment-ready assistant.
  • It represents a single seed experiment, with a multi-seed confirmatory replication in progress.
  • The emitted confidence scalar is noted as being close to constant in this lineage and should not be interpreted as fully calibrated.

Finetunes
1 models