| --- |
| language: |
| - en |
| license: cc-by-nc-4.0 |
| tags: |
| - pharmacovigilance |
| - medical |
| - mistral |
| - qlora |
| - faers |
| - drug-safety |
| - adverse-events |
| base_model: mistralai/Mistral-7B-Instruct-v0.3 |
| --- |
| |
| # pv-biomistral-7b |
|
|
| A pharmacovigilance-specialised language model fine-tuned from |
| [Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) |
| on 100,000 FAERS-derived training examples across five structured PV tasks. |
|
|
| This is the community testing release. It contains only the Q4_K_M quantized |
| GGUF for local inference via Ollama or llama-cpp-python. |
|
|
| --- |
|
|
| ## ⚠️ Important Disclaimer |
|
|
| This model is a **research prototype** intended for pharmacovigilance |
| professionals to evaluate and provide feedback on. It is **not a validated |
| system** and must not be used for: |
|
|
| - Autonomous pharmacovigilance decision-making |
| - Generating or contributing to regulatory submissions |
| - Replacing qualified pharmacovigilance assessor judgment |
| - Clinical or safety-critical decisions of any kind |
|
|
| All model outputs require review by a qualified pharmacovigilance professional. |
| This tool is for exploratory and research purposes only. |
|
|
| --- |
|
|
| ## Model Details |
|
|
| | Property | Value | |
| |---|---| |
| | Base model | mistralai/Mistral-7B-Instruct-v0.3 | |
| | Fine-tuning method | QLoRA (4-bit NF4, LoRA r=16) | |
| | Training records | 100,000 | |
| | Training epochs | 3 | |
| | Data source | FAERS public database (FDA) | |
| | Quantization | Q4_K_M (GGUF) | |
| | Model size | 4.37 GB | |
| | Context window | 8192 tokens | |
| | Framework | TRL 1.0.0, Transformers, PEFT | |
|
|
|
|
| ## Setup — Ollama (Recommended) |
|
|
| ### Requirements |
| - [Ollama](https://ollama.com/download) installed |
| - ~5 GB free disk space |
| - 8 GB RAM minimum, 16 GB recommended |
| - GPU optional but recommended for faster inference |
|
|
| ### Installation |
|
|
| **Step 1 — Download both files from this repository:** |
| - `pv-biomistral-7b-Q4_K_M.gguf` (4.37 GB) |
| - `Modelfile` |
|
|
| Place both in the same folder. |
|
|
| **Step 2 — Create the Ollama model** |
| ```bash |
| cd /path/to/downloaded/files |
| ollama create pv-mistral-v2 -f Modelfile |
| ``` |
|
|
| **Step 3 — Run** |
| ```bash |
| ollama run pv-mistral-v2 |
| ``` |
|
|
| **Windows users:** Use the full path e.g. `cd C:\Users\YourName\Downloads\pv-model\` |
|
|
| --- |
|
|
| ## Setup — llama-cpp-python (Alternative) |
|
|
| ```bash |
| pip install llama-cpp-python[server] |
| |
| python -m llama_cpp.server \ |
| --model pv-biomistral-7b-Q4_K_M.gguf \ |
| --chat_format mistral-instruct \ |
| --n_gpu_layers -1 \ |
| --n_ctx 8192 |
| ``` |
|
|
| Then open `http://localhost:8000/docs` for the Swagger UI. |
|
|
| --- |
|
|
| ## Setup — Jan App (Windows/Mac) |
|
|
| 1. Download [Jan](https://jan.ai) |
| 2. Import Model → select the GGUF file |
| 3. Set temperature to 0.1 in chat settings |
| 4. Add system prompt from the Modelfile SYSTEM field |
|
|
| --- |
|
|
| ## Expected Performance by Hardware |
|
|
| | Hardware | Speed | Response Time | |
| |---|---|---| |
| | Mac Mini M4 / Apple Silicon | 25-35 tokens/sec | 2-5 sec/case | |
| | Windows + NVIDIA GPU (8GB+ VRAM) | 25-40 tokens/sec | 2-4 sec/case | |
| | Snapdragon X Elite (16GB) | 8-15 tokens/sec | 5-12 sec/case | |
| | Windows CPU only (16-24GB RAM) | 3-6 tokens/sec | 15-30 sec/case | |
|
|
| --- |
|
|
| ## Known Limitations |
|
|
| - **Probable causality underrepresented:** Training data contained only 70 Probable |
| causality examples out of 100,000 records, reflecting real-world FAERS spontaneous |
| reporting patterns. The model may default to Possible even for cases with confirmed |
| positive dechallenge and no confounders. |
|
|
| - **Spontaneous reports only:** Trained exclusively on FAERS spontaneous adverse |
| event reports. Performance on clinical trial safety data, EHR-derived cases, |
| or non-English source material is untested. |
|
|
| - **Not formally validated:** The model has not been validated against any regulatory |
| standard including ICH E2D, ICH E2A, or WHO-UMC guidelines. |
|
|
| - **Short context optimised:** Designed for single-case inputs under 512 tokens. |
|
|
| --- |
|
|
| ## CIOMS WG XIV Alignment |
|
|
| This model is designed to operate within a Human-in-the-Loop (HITL) framework |
| consistent with CIOMS Working Group XIV recommendations for AI in drug safety. |
| All outputs are decision-support signals requiring human adjudication by a |
| qualified pharmacovigilance professional. |
|
|
| --- |
|
|
| ## Feedback |
|
|
| This is a community testing release. Please evaluate the model on real cases |
| from your practice area and share findings. Particular interest in: |
|
|
| - Causality outputs where you would classify Probable |
| - Cases with unusual drug combinations or rare reactions |
| - Narrative quality from a safety database entry perspective |
| - Therapeutic areas where performance appears weaker |
|
|
| --- |
|
|
| ## Training Data |
|
|
| Trained on 10,000 cases from the FDA Adverse Event Reporting System (FAERS), |
| accessed via public database export. No proprietary, confidential, or |
| patient-identifiable data beyond what is publicly available in FAERS was used. |
|
|
| --- |
|
|
| ## License |
|
|
| Base model (Mistral-7B-Instruct-v0.3): Apache 2.0 |
| Fine-tuned weights: CC BY-NC 4.0 (non-commercial research use only) |
|
|
| By downloading this model you agree to use it for research purposes only |
| and not for any commercial application or regulatory submission. |
|
|
