Open model release LoRA / PEFT adapter GGUF available

Qwen3-1.7B ResearchReasoning JSON RL

A compact adapter for evidence-grounded research question answering, source selection, claim verification, uncertainty reporting, and strict JSON schema generation.

Base modelQwen3-1.7B

AdapterLoRA / PEFT

Fine-tuning context4,096 tokens

LicenseApache-2.0

Diagnostic notice

Reported benchmark numbers are fixed-seed diagnostic runs using the project’s research JSON schema, not official leaderboard submissions.

Open the adapter View GGUF formats Review source code

Model overview

A structured research state, not only a free-form answer

Qwen3-1.7B ResearchReasoning JSON RL is an experimental adapter built on Qwen3-1.7B. It specializes the compact base model for source-aware question answering, evidence selection, claim verification, conflict detection, uncertainty handling, and schema-constrained JSON output.

The goal is to make the model expose its research plan, selected evidence, claims, conflicts, uncertainties, and final answer in a machine-checkable form.

Intended behavior

Evidence-grounded reasoning with explicit structure

Plan before answering.

Produces an explicit research plan and identifies the evidence needed for the task.

Select supplied sources.

Maps claims to source identifiers included in the prompt rather than inventing external citations.

Verify and qualify claims.

Reports supported, contradicted, or uncertain claims with confidence and evidence references.

Return strict JSON.

Targets a fixed schema suitable for validation, repair, and downstream research workflows.

Public benchmark diagnostics

Task specialization, not a general leaderboard claim

HotpotQA: 27.00 EM / 36.91 F1 on a 200-example fixed-seed diagnostic
PubMedQA: 52.50% label accuracy on an 80-example PQA-L diagnostic
SciFact: The adapter improves schema compliance and verifier reward but does not consistently improve classification over the base model
JSON validity: 95.50% on the reported HotpotQA diagnostic

These prompts require the project’s strict research JSON schema. Results are therefore not directly comparable with standard zero-shot leaderboard numbers.

Training design

Schema-focused adaptation with verifier-guided filtering

Base: Qwen/Qwen3-1.7B
Rank / alpha: LoRA rank 32, alpha 64, dropout 0.05
Final stage: One epoch over verifier-filtered rejection-sampling data initialized from a schema-hard SFT adapter
RL-lite: Verifier-guided rejection sampling followed by supervised fine-tuning; not a full online policy-gradient run
Languages: English and Turkish are represented in the project

Open research

Adapter, local formats, and reproducibility code

LoRA / PEFT adapter and model cardHugging Face · Adapter Quantized formats for local inferenceHugging Face · GGUF Training, evaluation, schemas, and reproducibility filesGitHub · QA-Research-SLM

Scope

Limitations

This is an experimental research adapter, not a production fact-checking system.
Generated evidence references are source IDs supplied in the prompt; the model does not browse or verify external URLs.
High verifier reward does not guarantee factual correctness.
Structured JSON can still be malformed or truncated.
Lower-bit GGUF quantization may reduce schema adherence.
Diagnostic results should be independently reproduced before comparison.

Reuse and citation

Cite the model release

@software{ethosoft_research_reasoner_2026,
  title      = {Qwen3-1.7B ResearchReasoning JSON RL},
  author     = {Ethosoft and QA-Research-SLM contributors},
  year       = {2026},
  url        = {https://huggingface.co/Ethosoft/Qwen3-1.7B-ResearchReasoning-JSON-RL},
  base_model = {Qwen/Qwen3-1.7B}
}