SmolLM3-Instruct Discharge Sentences SFT

A fine-tuned version of SmolLM3-3B for multi-label clinical sentence classification in hospital discharge summaries.

Model Description

This model classifies individual sentences from discharge summaries into categories of follow-up actions:

instructions: Case-specific instructions for the patient
appointment: Appointment-related followup
medication: Medication-related followups
lab: Lab-related followup
procedure: Procedure-related followup
imaging: Imaging-related followup
other: Other helpful contextual information

Performance

Evaluated on 5,313 test samples:

Metric	Score
JSON Validity	100.0%
Exact Match Accuracy	85.6%
Micro F1	0.796
Macro F1	0.647
Micro Precision	0.854
Micro Recall	0.745

Per-Category Performance

Category	Precision	Recall	F1	Support
instructions	0.873	0.745	0.804	1153
appointment	0.848	0.902	0.874	660
medication	0.797	0.674	0.730	239
lab	0.909	0.606	0.727	132
procedure	0.700	0.400	0.509	35
imaging	0.909	0.541	0.678	37
other	0.526	0.127	0.204	79

Comparison with Base SFT Model

Metric	Base SFT	Instruct SFT	Improvement
JSON Validity	97.5%	100.0%	+2.5 pp
Exact Match	76.3%	85.6%	+9.3 pp
Micro F1	0.631	0.796	+26%
Macro F1	0.568	0.647	+14%
Micro Recall	0.553	0.745	+35%

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model and adapter
base_model = AutoModelForCausalLM.from_pretrained(
    "HuggingFaceTB/SmolLM3-3B",
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "chrisvoncsefalvay/smollm3-instruct-discharge-sentences-sft")
tokenizer = AutoTokenizer.from_pretrained("chrisvoncsefalvay/smollm3-instruct-discharge-sentences-sft")

# Prepare input
messages = [
    {"role": "system", "content": "You are a clinical action item classifier..."},
    {"role": "user", "content": "Classify this sentence:\n\nPlease take aspirin 81mg daily."}
]

input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

# Generate
outputs = model.generate(**inputs, max_new_tokens=64, temperature=0.1)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)
# {"categories": ["medication", "instructions"]}

Training Details

Base Model: HuggingFaceTB/SmolLM3-3B (instruction-tuned)
Method: LoRA (r=64, alpha=128, dropout=0.1)
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Dataset: chrisvoncsefalvay/smol-discharge-sentences-sft
Training Samples: 25,782
Epochs: 3
Learning Rate: 5e-5
Effective Batch Size: 16
Precision: bf16

Advantages Over Base Model

Using the instruction-tuned SmolLM3-3B as a base provides:

Perfect JSON validity (100%) - built-in chat template ensures reliable structured output
Significantly higher recall (+35%) - better at identifying action items
No custom template needed - uses standard chat format

Citation

If you use this model, please cite:

@misc{smollm3-instruct-discharge-sft,
  author = {von Csefalvay, Chris},
  title = {SmolLM3-Instruct Discharge Sentences SFT},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/chrisvoncsefalvay/smollm3-instruct-discharge-sentences-sft}
}

License

Apache 2.0

chrisvoncsefalvay/smollm3-instruct-discharge-sentences-sft