dineshpotla/healthx-qwen35-4b-router-v1.8-experimental

HealthX Qwen3.5-4B Query Router (Experimental)

This repository contains the LoRA adapter only for the HealthX natural-language query router. It is not a standalone base model and does not contain the Qwen base weights. The adapter proposes a compact, schema-constrained routing decision (intent, optional entity and time scope, limit, and clarification state) for the HealthX read-only interoperability system.

Status and intended use

This is an experimental research artifact. It is not approved for production, clinical, diagnostic, prognostic, treatment, prescribing, or clinical-decision use. It must remain behind deterministic safety checks, patient/scope authorization, record-grounded entity resolution, strict schema validation, deterministic policy compilation, and read-only retrieval tools. The adapter cannot write to an EHR or select tools by itself.

The adapter passed the 142-example guarded recovery evaluation, but its one permitted disjoint external successor evaluation missed the release gates (99.3% schema validity and 87.5% unsupported-request recall against 99.5% and 95% minimums). The adapter therefore remains disabled by default in HealthX and must be explicitly opted into for evaluation.

Base model

• —Base: `Qwen/Qwen3.5-4B`
• —Pinned base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
• —Training method: 4-bit CUDA QLoRA with PEFT
• —LoRA rank / alpha / dropout: 16 / 32 / 0.05
• —Target modules: all-linear
• —Maximum sequence length: 512
• —Training steps: 600
• —Seed: 20260811

Training data

The router was trained on 4,720 balanced examples selected from a larger corpus of synthetic, patient-disjoint queries derived from synthetic FHIR concepts. The training and validation hashes, dependency versions, and adapter file hashes are preserved in training_manifest.json. No PHI, clinical databases, raw FHIR/C-CDA documents, Open-i images, or source tokens are included in this repository.

Files

• —adapter/adapter_model.safetensors: LoRA weights
• —adapter/adapter_config.json: PEFT configuration and base-model reference
• —adapter/tokenizer.json, adapter/tokenizer_config.json: tokenizer files
• —adapter/chat_template.jinja: chat template used by the evaluator
• —adapter/router_system_instruction.txt: constrained router instruction
• —training_manifest.json: reproducibility metadata and SHA-256 inventory

Loading with Transformers and PEFT

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.5-4B"
adapter_id = "dineshpotla/healthx-qwen35-4b-router-v1.8-experimental"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, subfolder="adapter")
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id, subfolder="adapter")

For the full HealthX integrity and safety path, download this repository and preserve its training_manifest.json plus the adapter/ directory as a local artifact. HealthX verifies the manifest hashes before loading an experimental adapter and requires an explicit experimental opt-in.

Limitations and safety

• —Synthetic-query evaluation does not establish clinical validity or real-user generalization.
• —The learned output is only a proposal; trusted deterministic code owns tool selection, authorization, retrieval, evidence composition, and abstention.
• —Invalid, unsafe, unsupported, or unauthorized requests must be rejected or routed to the deterministic fallback.
• —Never use this artifact to infer a diagnosis, treatment, patient identity, authorization, or clinical truth from missing data.

Licensing and attribution

This adapter is published as an experimental research artifact while its standalone redistribution license is under review. Users must comply with the license and terms of the `Qwen/Qwen3.5-4B` base model (listed there as Apache-2.0) and retain attribution for the synthetic Synthea training sources. This repository contains no patient-identifiable data.

Source project: dineshpotla/agentic-health-exchange