ade-extract-3b-v3

Fine-tuned locally with SLM Forge on 2026-09-25.

Goal: Rebuild ADE Extract as a consistent JSON-only information extractor: report only facts explicitly stated in the text, use null for missing optional details, and return an empty events list when no adverse effect is stated. Train a clean adapter from the original base; treat the result as an experimental extraction aid, not medical advice, and do not claim commercial clearance for the ADE Corpus v2 data unless its license is verified. Base model: mlx-community/Qwen2.5-3B-Instruct-4bit Runs on: Apple Silicon Macs with 8 GB unified memory or more.

Lineage

1. SFT (job 50): train_loss 0.018, peak_mem_gb 7.15, val_loss_start 0.875, val_loss 0.025

Usage

pip install mlx-lm
mlx_lm.generate --model engagepy/ade-extract-3b-v3 --prompt "Hello"

Recommended sampling

{
  "temperature": 0.7,
  "top_p": 0.95
}

System prompt

Extract only what the text explicitly states. Return only valid JSON in exactly this shape: {"events":[{"drug":null,"adverse_effect":"...","evidence_span":"...","severity":null,"dosage":null,"action_taken":null}]}. Include one event per explicitly described adverse effect. Fill drug, adverse_effect, severity, dosage, and action_taken only when each is stated in the text; otherwise use JSON null. Do not infer a drug-event link from timing alone, or infer causality, severity, dosage, or actions. Quote the shortest exact, contiguous evidence span supporting the event. If no adverse effect is stated, return {"events":[]}. This is an extraction aid, not medical advice.

Run it anywhere (GGUF)

GGUF builds of this model, for llama.cpp, Ollama and LM Studio:

• ade-extract-3b-v3-Q4_K_M.gguf (Q4_K_M, 1.80 GB)
• ade-extract-3b-v3-Q8_0.gguf (Q8_0, 3.06 GB)

llama-cli -m ade-extract-3b-v3-Q4_K_M.gguf -p "Hello"   # llama.cpp
ollama run hf.co/engagepy/ade-extract-3b-v3:Q4_K_M   # Ollama

In LM Studio, search for the repository or load the .gguf file.