occ-ai/OCC-RAG-1.7B-GGUF

OCC-RAG-1.7B-GGUF

<p align="center"> <img src="figures/occ.png" alt="OCC-RAG" width="320"/> </p>

<p align="center"> <a href="https://github.com/optimal-cognitive-core/OCC-RAG"><img src="figures/github-mark.png" alt=""/>&nbsp;<b>GitHub</b></a> &nbsp;|&nbsp; <a href="https://arxiv.org/abs/2606.00683"><b>Technical Report</b></a> &nbsp;|&nbsp; <a href="https://cloud.ru/products/evolution-ml-inference"><b>Cloud</b></a> &nbsp;|&nbsp; <a href="https://huggingface.co/occ-ai/OCC-RAG-1.7B"><b>Base model</b></a> </p>

GGUF quantizations of [`occ-ai/OCC-RAG-1.7B`](https://huggingface.co/occ-ai/OCC-RAG-1.7B) for native inference with llama.cpp, Ollama, LM Studio, and other GGUF-compatible runtimes.

OCC-RAG-1.7B is a 1.7B-parameter small language model specialized for faithful, context-grounded question answering: given a question and a set of sources, it produces a structured reasoning trace with explicit source citations, decides whether the context supports an answer, and either answers from the context or abstains. It attains the best faithfulness across all evaluated scales (up to 32B). See the base model card for training details and benchmarks.

Files

The underlying architecture is Qwen3 (1.7B). The chat template is embedded in the GGUF, so llama.cpp/Ollama apply it automatically.

For most uses pick Q4_K_M (smallest good quality) or Q8_0 (highest quality at under 2 GB). All quants are derived from the BF16 base. For an even smaller footprint or in-browser use, see `occ-ai/OCC-RAG-0.6B-GGUF`.

Usage — llama.cpp

# Run directly from the Hub (downloads the chosen quant)
llama-cli -hf occ-ai/OCC-RAG-1.7B-GGUF:Q4_K_M -p "Hello" -no-cnv

# Or download a file and run it
llama-cli -m OCC-RAG-1.7B-Q4_K_M.gguf -p "Hello" -no-cnv
# (newer llama.cpp: use `llama-completion` for non-interactive runs)

Usage — Ollama

ollama run hf.co/occ-ai/OCC-RAG-1.7B-GGUF:Q4_K_M

Input / output format

OCC-RAG uses a structured RAG prompt with special tokens: the question is wrapped in <|query_start|> … <|query_end|> and each source in <|source_start|><|source_id|>N … <|source_end|>. The response has five sections — query analysis → source analysis → reasoning → status (`ANSWERABLE` / `UNANSWERABLE`) → answer — and the final answer is in <|answer_start|> … <|answer_end|>.

The embedded chat template (apply with llama.cpp's --jinja) builds the query/source tokens for you when sources are supplied as documents; alternatively assemble the tokens manually. See the base model card for the full format and a runnable example.

We recommend greedy decoding (--temp 0), the training/evaluation default.

Limitations

• —Context-grounded only. Trained to answer from the supplied sources and to ignore parametric knowledge — not a general-purpose chat or knowledge model.
• —Reasoning depth. Training/evaluation are capped at three-hop reasoning; longer chains are out of distribution.
• —Quantization. Lower-bit quants (Q4) trade some quality for size; prefer Q6K/Q80 when accuracy matters most.

License

Released under the MIT License, inherited from the base model.

Citation

@misc{savkin2026occragoptimalcognitivecore,
  title         = {OCC-RAG: Optimal Cognitive Core for Faithful Question Answering},
  author        = {Maksim Savkin and Mikhail Goncharov and Alexander Gambashidze and Alla Chepurova and Dmitrii Tarasov and Nikita Andriianov and Daria Pugacheva and Vasily Konovalov and Andrey Galichin and Ivan Oseledets},
  year          = {2026},
  eprint        = {2606.00683},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2606.00683}
}