organ-lm — an LLM with a native, calibrated numerical organ

organ-lm grafts a 3.4M-parameter tabular estimator — trained from scratch on a purely synthetic prior, zero real data — into a frozen Qwen3-1.7B as 8 soft tokens. The LLM never sees the organ's digits; it reads the organ's full posterior (a 256-bin bar distribution) through a trained projection and answers numeric questions with the organ's calibration. A bypass serving mode splices the organ's decoded values verbatim into the LLM's answer structure — numbers the model cannot hallucinate.

Code, demo, and full results with caveats: https://github.com/matej-01RAI/organ-lm

Results (summary)

• Three internally pre-registered gates, all PASS: the graft beats the strongest text-serialization baseline on point + distributional metrics (p ≤ 0.0005), preserves the organ's 90%-interval calibration within 0.05 coverage, and leaves language ability intact (20/20 sanity prompts).
• On identical 256-row evaluation cells, the organ system leads every non-organ baseline on calibration (coverage 0.914 vs nominal 0.90; TabPFN 0.919 distance-wise behind, CatBoost 0.61, plain LLM 0.07) and sits second on point accuracy behind TabPFN, ahead of untuned CatBoost. Small-context regime; see the repo's RESULTS.md for honest scoping — this is not a tabular-SOTA claim.
• Bypass router: 0 misses in 6,400 episodes; bypass output is exactly organ-identical.

Files

path	what
graft/projector.pt	payload→soft-token projector (~17M params)
graft/graft_meta.json	graft config (k_soft=8, dims)
graft/adapter/	LoRA adapter (r=32) for Qwen/Qwen3-1.7B
organ/plt_r5.pt	the 3.4M organ checkpoint (from our internal PLT research program)

Use

Clone the GitHub repo and place: graft/ → demo/artifacts/graft_best/, organ/plt_r5.pt → demo/artifacts/plt_r5_ckpt.pt, then demo/run_demo.sh (interactive local demo; Apple silicon or CPU). The organ's model code ships vendored in the repo — no extra dependencies beyond torch transformers peft gradio numpy matplotlib.

Limitations

Proof-of-concept: the graft's interface is format-locked to its JSON training schema; organ context is capped at 512 rows / 64 features; evaluation is a 20-dataset internal suite (OpenML CC18/CTR23 subsets), 3 seeds. Base model © Alibaba/Qwen team, Apache-2.0, downloaded separately — not redistributed here.

Zero One Research, 2026. Apache-2.0.