liuiop/CDIDNMA

CAIAAD — evidence-grounded scientific fact extraction (LoRA adapter)

LoRA adapter for Qwen/Qwen3-VL-8B-Instruct that turns paper passages into evidence-grounded structured facts: a verbatim quote, a complete numeric value, the original unit and a field from a fixed 26-field taxonomy.

Trained for the CDI / water-treatment corpus of this project (capacitive deionization, electrosorption, antibacterial surfaces, ROS/H2O2).

Usage

from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer

base = AutoModelForImageTextToText.from_pretrained(
    "Qwen/Qwen3-VL-8B-Instruct", dtype="bfloat16", device_map={"": 0})
model = PeftModel.from_pretrained(base, "liuiop/CAIAAD")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-VL-8B-Instruct")

The system prompt used during training is reproduced in training-config.json of the training run; the adapter expects chat messages of the form system=<extraction protocol>, user={"task": "structured_scientific_extraction", "passages": [<chunk>]}.

Training data

• 7,556 facts over 1,974 passages from 955 publications.
• Targets are validator-audited Qwen3-VL-30B extractions, not human annotation. Every fact already passed a deterministic validator (verbatim quote, value inside the quote, allowed field, unit adjacency, assay/attribution rules).
• Splits are made by publication DOI, so no paper appears in two splits.

split	passages	facts	publications
train	1,587	6,049	779
validation	199	776	82
test	188	731	94

Training setup

item	value
method	supervised fine-tuning (SFT) with LoRA, causal-LM loss on assistant tokens only
precision	bf16, no quantization
LoRA	r=16, alpha=32, dropout=0.05, q/k/v/o/gate/up/down projections
trainable parameters	43,646,976 (0.50% of the base model)
optimizer	AdamW (betas 0.9/0.95, weight decay 0.01), OneCycleLR
batching	batch 1 × gradient accumulation 8, sequence 1,536
steps	796 (epoch 4 of 8 trained)
hardware	single RTX 5090, 24 GB, 22.0 GB peak

Evaluation

Held-out extraction on 12 unseen passages, every produced fact re-checked by the same deterministic validator:

checkpoint	facts produced	passed validator	precision
this adapter (epoch 4)	44	37	84%
epoch 8 (final)	43	32	74%
base Qwen3-VL-8B (no adapter)	50	0	0%

The base model does not follow the output contract at all; this adapter does.

Limitations

• Targets are machine-generated and validator-audited, not human-adjudicated. Upstream audit removed 1,132 of 8,745 candidate facts (591 wrong claims, 541 pending rule review).
• Validation loss falls to 0.0897 at epoch 4 and rises afterwards (0.1025 at epoch 8); later checkpoints over-fit and were not published.
• Recall is uneven: some passages yield too few facts, others over-extract (one passage produced 10 facts where the reference set had 2).
• Trained for CDI / water-treatment literature; other domains are untested.
• The adapter may emit verbatim sentences from scientific papers. Cite the original publications when reusing extracted quotes.

License

Apache-2.0, matching the base model. Training data derive from published scientific literature; the adapter does not redistribute the corpus, and the training dataset is not included in this repository.