amkkk/Trace-Inverter-4B-NoBubble

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 21, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

amkkk/Trace-Inverter-4B-NoBubble is a 4-billion parameter trace inversion model based on Qwen3-4B-Instruct-2507, developed by amkkk. It is specifically trained to reconstruct detailed synthetic reasoning traces from an original problem and a known final answer, without requiring a reasoning bubble or summary at inference time. This model excels at generating plausible reasoning paths for historical logs or research into no-bubble trace inversion.

Loading preview...

Trace-Inverter-4B-NoBubble: No-Bubble Trace Inversion

Trace-Inverter-4B-NoBubble is a 4-billion parameter model developed by amkkk, built upon Qwen3-4B-Instruct-2507. Its core function is to reconstruct a detailed synthetic reasoning trace (<think>...</think>) given only an original problem and a final answer, crucially without requiring any reasoning bubble or summary during inference. This distinguishes it from traditional trace inversion models that rely on such intermediate reasoning.

Key Capabilities

  • Synthetic Trace Reconstruction: Generates plausible step-by-step reasoning that could lead to a given final answer.
  • No-Bubble Inference: Operates effectively in scenarios where intermediate reasoning (like Chain-of-Thought or summaries) is unavailable.
  • Distillation of Bubble-Assisted Inversion: Learns from targets originally produced by a bubble-conditioned inversion pipeline, but is trained to perform the task without the bubble input.
  • Format Discipline: Outputs traces consistently within <think> and </think> tags.

Good For

  • Research: Studying the no-summary setting of trace inversion and measuring the information carried by reasoning bubbles.
  • Historical Data Analysis: Reconstructing reasoning for historical chat logs or datasets where only questions and answers exist.
  • Synthetic Data Generation: Building synthetic reasoning-supervision data, with the caveat that downstream verification is mandatory.

Limitations

It's important to note that the generated traces are synthetic reconstructions and not authentic hidden reasoning. The model is conditioned on the supplied final answer and will construct a plausible path even to an incorrect answer, necessitating verification. It is not a general-purpose chat model or a verifier for maths/code.

Finetunes
2 models