Rumiii/Trion_8B-FHIR_R4
Rumiii/Trion_8B-FHIR_R4 is an 8 billion parameter Llama-3.1-8B-Instruct model fine-tuned by Rumiii for multi-turn clinical tool-calling over FHIR R4 APIs. This model specializes in planning and executing sequences of structured function calls for clinical instructions, such as patient lookup, lab queries, and medication ordering. It uniquely converts raw text-action formats into Llama 3.1's native JSON tool-calling chat template, making it a direct fit for standard Transformers pipelines. The model's training pipeline surgically compresses FHIR JSON responses and applies response-only loss masking to optimize for clinical reasoning within an 8192 token context length.
Loading preview...
Trion-8B-FHIR-R4: Specialized FHIR R4 Clinical Tool-Calling Model
Trion-8B-FHIR-R4 is a fine-tuned Llama-3.1-8B-Instruct model developed by Rumiii, specifically designed for multi-turn clinical tool-calling using FHIR R4 APIs. Unlike other models, it was trained to emit calls using Llama 3.1's native JSON tool-calling chat template, ensuring seamless integration with standard transformers pipelines.
Key Capabilities & Differentiators
- FHIR R4 Tool-Calling: Excels at planning and executing sequences of structured function calls for clinical tasks, including
search_patient,get_observation,record_vital_sign,order_medication,order_service_request, andfinish. - Optimized Training: The training pipeline significantly compresses FHIR JSON tool responses (e.g., from ~3,000 tokens to ~50) to focus the model's attention on clinical reasoning rather than boilerplate. It also converts raw
GET/POST/FINISHactions into proper OpenAI/Llama-styletool_calls. - Efficient Learning: Utilizes response-only loss masking, meaning the model is only trained to predict assistant tool-call tokens, ignoring system prompts, user instructions, or tool outputs it receives. This ensures zero truncation within its 3072-token training budget.
Intended Use Cases
- Clinical LLM Agent Prototyping: Ideal for research and development of clinical LLM agents that interact with FHIR-based EHR systems via tool-calling.
- Workflow Automation: A strong starting point for building assistants that can chain complex clinical workflows like patient lookup, lab retrieval, and order placement.
- Educational Reference: Serves as an excellent example for adapting small, 'flat' tool-calling datasets into efficient instruction-tuned chat models.
Limitations
It is not validated for clinical use and should not be used for real patient care. The model was trained on a limited dataset (284 examples, 10 task types) and has a fixed six-function tool schema, limiting its generalization to arbitrary or unseen tool definitions or uncompressed FHIR payloads.