Strands Decider 2B — Korean Clinical

Reads a doctor's shorthand. Answers in probabilities. Fits in 0.9 GB.

Korean outpatient notes are terse, bilingual and full of hedges: 물치 의뢰, f/u 2w, PT 거부함, 호전 없으면 MRI 고려. This model reads a plan line together with the patient summary and the doctor's own habit notes, and answers the questions that turn a note into action — Is this a referral? A follow-up? How many weeks? Which handout? Which past plan fits this patient? — each with a calibrated probability, so you can act only when it is sure. It never writes text: every answer is one of the options you supply.

Highlights

  • 🎯 70.1% → 97.6% on held-out lines and doctors after LoRA fine-tuning; the compressed release keeps 96.1%.
  • 📈 +9 points in a full autocomplete pipeline on 144 independently written cases (79.2% → 88.2%), with wrong suggestions nearly halved (20 → 11).
  • 🧭 Reads negation and context: refused, deferred, conditional and already-done plans, and picks the past plan that fits each patient (21.8% → 85.2%).
  • 🪶 4× smaller, same pipeline score: LoRA merged, 4 of 24 layers pruned, int4 weights. 3.8 GB → 0.9 GB; 127/144 before and after.
  • ⚡ 77 ms per request (9–11 questions) on a single RTX PRO 6000, 2× faster than serving the unmerged LoRA.
  • 📏 Trustworthy confidence: ECE 0.022; when it says ≥ 0.9, it is right 97.2% of the time.
  • 🔓 Apache 2.0, with the full training, pruning and quantization recipe below.

⚠️ Research model — not a medical device. Trained and evaluated only on synthetic data. Do not use for patient care.

Model details

Developed by William Park
Model type Strands Decider: Qwen3.5-2B torso + pointer head that scores options
Fine-tuned from StrandsAgents/strands-decider-2B-qwen3.5-v1-2610
Language Korean and English clinical shorthand (mixed)
License Apache 2.0
Architecture 20 decoder layers (Gated DeltaNet + full attention), 1.65B parameters (the size widget counts packed int4 bytes, so it shows ~0.8B)
Weights int4, group size 128, 0.9 GB on disk (3.8 GB before compression)

Uses

Intended use. Research on calibrated, option-scoring models for short clinical text: intent detection under negation and conditions (refused, considered, deferred, already done), slot filling from habit notes, and patient-aware selection.

Out of scope. Real patient data or clinical decisions; free-text generation; chart formats and languages outside the synthetic Korean outpatient format it was trained on.

Risks and limitations

  • Synthetic data only. Real chart language, abbreviations and edge cases will differ.
  • Generator bias. Accuracy on held-out rows from the training generator is far higher than on text written independently (see Evaluation).
  • Negation scope remains the hardest case (considered, other hospital, already done: 62–76% after training, on small held-out slices of 19–42 rows).
  • Probabilities are calibrated on synthetic validation data; recalibrate before using a confidence threshold on other data.

How to get started

The weights are stored as packed int4. load_tabchart.py (in this repo) unpacks them to bf16 for the Strands Decider server.

pip install torch==2.7.1 transformers==5.17.0 peft==0.21.0 safetensors huggingface_hub
git clone https://github.com/strands-labs/strands-decider && pip install -e strands-decider
python load_tabchart.py carlclausewitz/strands-decider-clinical-use ./decider
strands-decider serve ./decider/ckpt --port 8000

Requests follow the Strands Decider format (state text + questions + options). A GPU or an Apple-silicon machine with 16 GB+ memory is recommended.

Training details

Training data

Synthetic Korean outpatient data:

  • Plan lines written by five LLM writers that saw only a label specification (no test data): Korean-style lines, English-heavy lines, minimal pairs, keyword-free lines, and chart notes / habit-note templates. About 1,500 annotated lines, split by text into training and validation (no shared line, even with numbers changed).
  • Synthetic world: randomized virtual doctors (facility, frequency, duration, handout, follow-up interval, condition → department habits) and patients (diagnoses, eGFR, anticoagulants, NSAID allergy, red flags).
  • Rows: one (state, question, gold option). Rows whose gold answer is ambiguous under the label specification are not created.
  • Audited twice by an independent LLM auditor; generator errors found in the first audit were fixed before training.
  • 56,725 training rows, 9,871 validation rows (160 training doctors, 30 separate validation doctors).

Training procedure

Hyperparameter Value
LoRA rank 48, alpha 96, dropout 0.05; attention, MLP and Gated DeltaNet projections
Trainable parameters 53.6M / 1.94B (2.77%)
Epochs · steps · batch 2 · 3,544 · 32
Learning rate LoRA 1e-4, head 3e-4, warmup 3%, weight decay 0.01
Regularization KL to the frozen base decider (weight 0.3), option shuffling
Calibration temperature scaling (ECE 0.182 → 0.015)

Validation loss fell from 0.213 (step 800) to 0.151 and stayed flat; KL to the base model stayed near 0.4 (0.35–0.45) without drifting.

Compression

  1. Merge the LoRA adapter into the base weights.
  2. Prune layers 9, 13, 18, 20: the least important linear-attention layer in each block, by single-layer skip accuracy (the first block and all full-attention layers were kept). Healing distillation was tested and not kept: it recovered row accuracy but added one wrong suggestion on the dev set.
  3. Quantize all linear layers and the tied embedding/lm_head to symmetric int4, group size 128.

Evaluation

Held-out validation rows (unseen lines, unseen doctors):

Question Base decider After LoRA (bf16)
Past-plan selection 21.8% 85.2%
Which habit note applies 45.9% 95.4%
Is it a follow-up? 48.4% 98.7%
Is an action committed now? 61.2% 95.9%
All rows (9,871) 70.1% 97.6%

Compression (3,430-row validation subset):

Model Layers Accuracy Acc. when conf ≥ 0.9 ECE
LoRA, bf16 24 96.8% 97.1% 0.024
+ 4 layers removed 20 96.3% 97.2% 0.021
+ int4 g128 (this model) 20 96.1% 97.2% 0.022
int4 per-channel (not used) 24 95.0% 95.6% 0.041
int3 g128 (not used) 24 81.3% 89.8% 0.113

On 144 independently written cases scored in a full autocomplete pipeline (rule parser + this model + confidence threshold), accuracy rose from 79.2% (base) to 88.2% (LoRA), and the compressed model matched the uncompressed LoRA exactly (127/144, 11 wrong, 6 missed).

Speed: median 0.077 s per request of 9–11 questions on an RTX PRO 6000 (0.152 s for the unmerged LoRA). int4 is a storage format here; weights are unpacked to bf16 at load time.

Compute

NVIDIA Brev cloud GPUs (RTX PRO 6000, L4): about 7.7 GPU-hours for training, compression and evaluation (≈ US$20).

Citation

@misc{park2026decider,
  title  = {Strands Decider 2B for Korean Clinical Plan Lines: LoRA, Pruning and int4 Quantization},
  author = {Park, William},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/carlclausewitz/strands-decider-clinical-use}}
}

Contact

William Park — open a discussion in the Community tab.

Downloads last month
48
Safetensors
Model size
0.8B params
Tensor type
F32
·
BF16
·
U8
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for carlclausewitz/strands-decider-clinical-use

Adapter
(49)
this model