Trace-Inverter-4B-NoBubble-GGUF

Trace-Inverter-4B-NoBubble GGUF

This is the GGUF Q8_0 release of amkkk/Trace-Inverter-4B-NoBubble. It contains the same merged Trace-Inverter-4B-NoBubble weights converted for llama.cpp-compatible runtimes.

Trace-Inverter-4B-NoBubble is a 4B-parameter no-bubble trace inversion model. Given an original problem or conversation context and a known final answer, it reconstructs a detailed synthetic reasoning trace wrapped in <think> and </think>. No reasoning bubble, reasoning summary, scratchpad, or hidden chain-of-thought is required as input.

Generated traces are synthetic reconstructions. They must not be presented as the actual hidden reasoning of Claude, Qwen, or any source model.

File

• Trace-Inverter-4B-NoBubble-Q8_0.gguf - GGUF Q8_0 quantization

The GGUF preserves the model tokenizer and chat template metadata from the BF16 Transformers release.

Base Model

Qwen/Qwen3-4B-Instruct-2507

Relationship To BF16 Release

This GGUF is a quantized release of the same weights hosted at amkkk/Trace-Inverter-4B-NoBubble. See the BF16 model card for full training details, dataset provenance, limitations, and benchmark methodology.

This is the NoBubble inverter. It is not Jackrong/Trace-Inverter-4B and does not consume reasoning bubbles at inference time.

Usage

With llama.cpp:

llama-cli -m Trace-Inverter-4B-NoBubble-Q8_0.gguf -p "<your chat-formatted prompt>" -n 4096

With Ollama, create a Modelfile next to the GGUF:

FROM Trace-Inverter-4B-NoBubble-Q8_0.gguf
PARAMETER temperature 0
PARAMETER num_predict 4096

Then run:

ollama create trace-inverter-4b-nobubble-gguf -f Modelfile
ollama run trace-inverter-4b-nobubble-gguf

Recommended prompt shape:

Problem:
{problem}

Model's final answer:
{final_answer}

Reconstruct the detailed synthetic reasoning trace.

Expected output:

<think>
...synthetic reconstructed reasoning...
</think>

Public-10 GGUF Check

Deterministic public-10 evaluation was run locally through llama.cpp server against data/processed/public10.jsonl.

License

Apache 2.0. The BF16 release, base model, and source datasets are Apache 2.0.