Onca 1.0 9B Int4
Model Summary
This repository contains the 4-bit release of Onca 1.0, an open 9B pancreatic cancer language model specialized for:
- clinical trial screening
- case-specific clinical reasoning
- structured pathology report extraction
- molecular variant evidence reasoning
This variant is intended for more memory-constrained inference setups that still need access to the same Onca 1.0 task scope.
Base Model
Onca 1.0 is fine-tuned from Jackrong/Qwopus3.5-9B-v3, a Qwen3.5-derived 9B dense reasoning model. This repository packages the 4-bit quantized Onca 1.0 checkpoint with a bitsandbytes NF4 configuration.
Variant Details
- Quantization: 4-bit
- Quantization method:
bitsandbytes
- Quantization type: NF4 with double quantization
- Recommended use: memory-constrained inference in compatible
transformers environments
For the main reference release, use JosephKBS/onca-1.0-9B.
Training Scope
The underlying model was trained on 37,364 prepared rows from openly available sources across four pancreatic-cancer task families:
- trial eligibility screening
- oncology clinical reasoning
- CAP-aligned pathology abstraction
- CIViC-style variant interpretation
Intended Use
This release is intended for research, benchmarking, and local experimentation in oncology-focused NLP workflows where lower memory usage matters.
Out-of-Scope Use
This model is not intended for direct clinical care, autonomous decision-making, or unsupervised patient-facing use. It is a research checkpoint and does not replace clinician judgment.
Evaluation Summary
The evaluation reported for Onca 1.0 in the companion manuscript includes:
- Trial Screening: 81.6 F1
- Clinical Reasoning: 14.1 composite
- Pathology Extraction: 30.5 field exact-match
- PubMedQA Cancer: 68.3 macro-F1
- PubMedQA: 66.5 macro-F1
These values summarize the Onca 1.0 model family and should not be read as a separate benchmark guarantee for every hardware/runtime combination used with the 4-bit checkpoint.
Limitations
- Specialized for PDAC and adjacent oncology tasks rather than general medicine
- Public-source training data improve reproducibility but do not fully capture institutional documentation style
- Quantized inference behavior depends on runtime support, kernels, and available hardware
- Research-only checkpoint; not clinically validated
Usage
This repository includes a bitsandbytes 4-bit configuration in config.json. In a compatible environment, a standard loading pattern is:
1from transformers import AutoTokenizer, AutoModelForCausalLM
2
3model_id = "Joesh1/onca-1.0-9B-Int4"
4
5tokenizer = AutoTokenizer.from_pretrained(model_id)
6model = AutoModelForCausalLM.from_pretrained(
7 model_id,
8 device_map="auto",
9)
Please ensure your local environment supports bitsandbytes 4-bit inference before deployment.
Files in This Repository
model.safetensors: quantized model weightsconfig.json: architecture and quantization configurationgeneration_config.json: default generation settingstokenizer.json and tokenizer_config.json: tokenizer fileschat_template.jinja: chat formatting template
License
This release is provided under the Apache 2.0 license. Users should also review the upstream base-model terms and the terms of any underlying datasets or benchmarks referenced by the project.
Citation
If you use Onca 1.0, please cite the accompanying manuscript when publicly available. A temporary reference is:
1@misc{shim2026onca,
2 title = {Onca: An Open 9B Language Model for Pancreatic Cancer Clinical Tasks},
3 author = {Shim, Kwan Bo},
4 year = {2026},
5 note = {Preprint in preparation}
6}
Acknowledgments
This project builds on the work of the Qwen and Qwopus model developers, together with the public institutions, benchmarks, and open data contributors that made this release possible.
