MetaResearch PapersRAG 1.5B

One-line: Retrieval-augmented generation system for querying recent scientific literature with daily auto-updates from arXiv cs.CL.

Coding benchmarks

Benchmark	Score	Source
HumanEval	—	—
HumanEval+	—	—
MBPP	—	—
LiveCodeBench	—	—
Aider polyglot	—	—
SWE-bench Verified	—	—
SWE-bench Multilingual	—	—

Agentic / tool-use benchmarks

Benchmark	Score	Source
BFCL	—	—
τ-Bench	—	—
ToolACE	—	—
GAIA	—	—

Reasoning (context only)

Benchmark	Score	Source
MMLU / MMLU-Pro	—	—
GPQA Diamond	—	—
MATH	—	—
AIME 2025	—	—

Efficiency

• Active params: 1.5B (Qwen2.5-1.5B base)
• Context length: 2048 tokens (base model default)
• VRAM (fp16 rough): ~3 GB (1.5B params × 2 bytes)
• Architecture: Dense transformer with RAG pipeline (bi-encoder retrieval + cross-encoder re-ranking + LM generation)

What makes it notable

PapersRAG-1.5B is a specialized RAG system rather than a general-purpose coding model. It pairs Qwen2.5-1.5B with a FAISS index of arXiv cs.CL paper abstracts, updated daily via automated pipeline. The system uses dense retrieval + cross-encoder re-ranking to provide citation-backed answers. While not designed for coding tasks, it could serve as a research assistant component in agentic loops that need scientific literature context. The Apache-2.0 license and small size make it suitable for local deployment. However, it lacks coding/agentic benchmarks and is limited to English-language cs.CL papers only.

See also

• welcome
• Source: https://huggingface.co/metaresearch/PapersRAG-1.5B