Kuiper R1

Planetary Class · 9.41B · Apache 2.0 Trace inversion Synthetic reasoning Safetensors Public release Expand the trace. Preserve the answer.

Overview

Kuiper-R1 reconstructs full, readable reasoning from a problem, its final answer, and compressed reasoning bubbles.

Each result follows a three-tag contract: <synthetic_trace>, <verification_note>, and <final>. The final answer is preserved byte-for-byte, while the reconstructed trace is explicitly labeled synthetic.

Use Kuiper-R1 for reasoning-dataset augmentation, interpretability review, and research into reasoning compression and expansion. Do not present its synthetic traces as another model's hidden chain-of-thought.

Specifications

02 Architecture TypeQwen3.5 conditional generation Params9.41B Context262,144 tokens ArtifactBF16 safetensors Training MethodQLoRA SFT → ORPO → merge SFT rows3,000 Preference87 scored pairs DataOpenThoughts3-1.2M

Usage

Serve with vLLM. Disable the default thinking pass so the requested three-tag contract can use the full output budget.

Recommended Samplers Temperature0.4 Repetition1.1 ThinkingDisabled Model len8,192 served

Start the checkpoint locally:

vllm serve Michael-Kozu/Kuiper-R1 \
  --served-model-name kuiper-r1 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.85 \
  --skip-mm-profiling \
  --trust-remote-code

Send an OpenAI-compatible request:

{
  "model": "kuiper-r1",
  "messages": [
    {
      "role": "system",
      "content": "You are Kuiper, a reasoning-trace inversion model."
    },
    {
      "role": "user",
      "content": "Problem:\n[problem text]\n\nAnswer:\n[exact final answer]\n\nReasoning Bubbles:\n- [bubble one]\n- [bubble two]\n\nReconstruct the full reasoning trace."
    }
  ],
  "temperature": 0.4,
  "repetition_penalty": 1.1,
  "chat_template_kwargs": {"enable_thinking": false}
}

Training Details

Training uses domain-balanced science, math, and code rows from OpenThoughts3-1.2M. Reasoning traces are compressed into extractive bubbles deterministically, without an additional model call.

3,000SFT examples 1SFT epoch 87ORPO pairs

Training split: 1,531 science, 1,001 math, and 468 code examples.
Compression is extractive at a target ratio of 0.28 with 3–24 bubbles.
The held-out DeepSeek-V4 benchmark lineage is disjoint from the QwQ-32B training lineage.

Benchmarks

05 Kuiper-R1 release · held-out n=200 Answer exact 0.820 Format valid 0.880 Expansion 0.769 Overall 0.504

Metric	Base	Kuiper-R1
Final answer preserved · exact	0.155	0.820
Format valid · 3-tag contract	0.980	0.880
Expansion quality	0.791	0.769
Semantic consistency	0.708	0.535
No invented tool outputs	1.000	1.000
Overall	0.329	0.504

Limitations & License

Bubble informativeness bounds trace quality; terse inputs provide little evidence to expand.
The verification note is a construction-time consistency statement, not an external verifier.
The release improves exact answer preservation but trails the untouched base on raw format validity and semantic consistency.

The checkpoint is released under the Apache License 2.0.

Kozu AI Turning the laws of reality into unparalleled creation.