Qwen3-1.7B ResearchReasoning JSON RL — GGUF
GGUF quantizations of Ethosoft/Qwen3-1.7B-ResearchReasoning-JSON-RL, a Qwen3-1.7B fine-tune for evidence-grounded research reasoning, source-aware QA, claim verification, and strict JSON schema generation.
The PEFT adapter was merged into the base model before GGUF conversion. The project calls the final training stage RL-lite: verifier-guided rejection sampling followed by supervised fine-tuning, not full online policy-gradient RL.
Files
| Quantization | Suggested use |
|---|
Q8_0 | Highest fidelity among the uploaded quantizations; largest file |
Q5_K_M | Recommended balance for schema-sensitive local use |
Q4_K_M | Smaller and faster; validate JSON carefully |
GGUF structural validation
Each file was checked for a valid GGUF binary header, supported GGUF version, nonzero tensor count, and nonzero metadata count. The files were produced successfully with the llama.cpp converter and quantizer. This is a structural file validation, not a quant-specific inference or JSON benchmark.
| Quantization | Size | GGUF version | Tensors | Metadata entries | Structural check |
|---|
Q8_0 | 1.71 GiB | 3 | 310 | 29 | Passed |
Q5_K_M | 1.17 GiB | 3 | 310 | 29 | Passed |
Q4_K_M | 1.03 GiB | 3 | 310 | 29 | Passed |
Raw results are available in structural_validation.json.
llama.cpp
1llama-cli \
2 -m Qwen3-1.7B-ResearchReasoning-JSON-RL-Q5_K_M.gguf \
3 -cnv \
4 -n 768
For programmatic use, supply a strong system prompt that requires valid JSON, lists all required fields, disables markdown, and restricts evidence entries to source IDs present in the prompt.
Recommended output schema
1{
2 "task_type": "document_qa",
3 "research_plan": ["read the question", "inspect the sources", "answer with evidence"],
4 "evidence_needed": ["supporting source text"],
5 "selected_sources": ["src_001"],
6 "claims": [
7 {
8 "claim": "The answer is supported by src_001.",
9 "status": "supported",
10 "confidence": 0.9,
11 "evidence": ["src_001"],
12 "importance": "high"
13 }
14 ],
15 "conflicts": [],
16 "uncertainties": [],
17 "answer": "Evidence-grounded answer"
18}
Benchmark diagnostics
The full diagnostic results were produced with the PEFT adapter before GGUF conversion. They are included here for model context, but should not be treated as quant-specific measurements.
| Benchmark | Adapter result | Diagnostic size |
|---|
| PubMedQA PQA-L | 52.50% label accuracy | 80 |
| HotpotQA Distractor | 27.00 EM / 36.91 F1 | 200 |
| SciFact claims-dev | 40.00% label accuracy / 26.21 macro-F1 | 200 |
These are fixed-seed custom diagnostics using a strict JSON research prompt, not official leaderboard submissions. See benchmark_results.json and the adapter model card for full settings and base-model comparisons.
Quantization caution
This model is optimized for strict structured generation. Quantization can preserve semantic answers while still damaging commas, quotes, closing braces, field names, or source-ID arrays. Validate every response. For reliability-sensitive applications, prefer Q5_K_M or Q8_0 and use constrained decoding or application-level JSON repair.
VibeThinker inspiration and attribution
This project is conceptually inspired by the VibeThinker line of work on verifiable reasoning in compact models, but it was independently trained for research-oriented structured generation. It does not reuse VibeThinker weights, data, or code.
- Sen Xu et al., Tiny Model, Big Logic: Diversity-Driven Optimization Elicits Large-Model Reasoning Ability in VibeThinker-1.5B, arXiv:2511.06221, 2025. https://arxiv.org/abs/2511.06221
- Sen Xu et al., VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models, arXiv:2606.16140, 2026. https://arxiv.org/abs/2606.16140
License
Apache-2.0, following the Qwen3 base model license. Review the original dataset licenses before redistributing derived training data.
Quant-specific evaluation status
The GGUF files passed structural validation after conversion and quantization. A CPU-only llama.cpp generation attempt on the TRUBA node exceeded a five-minute startup timeout for the Q8_0 file, so this release does not claim completed quant-specific runtime, JSON-adherence, or public-benchmark results. The benchmark table above belongs to the PEFT adapter before GGUF conversion. Re-run the evaluation suite in your target llama.cpp, LM Studio, or Ollama environment before reliability-sensitive deployment.