Sold by: Empirisch Tech GmbH
Deployed on AWS
Free Trial
Drug-level sentiment for pharma and medical text. For each drug or therapy you name, Chaperone-SA-1.0 says whether a conference abstract, publication, trial record, news article or physician's post presents it as Positive, Neutral or Negative, as JSON. 91% accurate on 473 held-out test labels. It runs on one GPU in your own AWS account, network-isolated, as a real-time endpoint or a batch transform job.
Overview
Chaperone-SA-1.0 shows how pharma and medical text talks about each drug, product or therapy you track. Send it a document with up to 8 targets, and it answers Positive, Neutral or Negative for each, as a JSON object. It reads conference abstracts, PubMed publications, ClinicalTrials.gov records, physicians' social media posts and press articles, and is built for medical affairs, competitive intelligence, market access and commercial insight teams.
MEDICAL KNOWLEDGE NEXT TO FRONTIER MODELS
- MedQA, US medical licensing exam, thinking mode on: 94.2%, about 2 points below the top of the Vals.ai leaderboard (o1, 96.5%, April 2026).
- MedMCQA, 1,000-question sample: 81.4%. Best frontier score found: o1-preview, 83.9%.
- MMLU medical, step by step: 94.2%. GPT-5: 96.0%.
- PubMedQA, thinking mode on: 80.2%. GPT-4 with Medprompt: 82.0%. The knowledge comes from the base model, and fine-tuning kept it: 86.5% against 85.9% on 3,091 exam questions. We ran our model ourselves; frontier scores come from public sources with slightly different setups, so treat 1-2 points as a tie. Published frontier scores on MedMCQA and MMLU medical stop at o1-preview and GPT-5. Sentiment analysis runs with thinking mode off.
DRUG-LEVEL SENTIMENT, TESTED ON HELD-OUT DOCUMENTS
- 91% accurate: right on 432 of 473 held-out test labels (95% range 88-94%), and 87-95% on each of the five content types above.
- Finds 84% of negative mentions, and 84% of its negative calls are right (45 negatives; 95% range 71-92%).
- A valid label for every requested drug: 0 failures in 1,500 test answers.
- 8 to 15 points ahead of the other open-source models we tested on the same labels: Qwen3.8-27B 83.5%, Qwen3-32B 83.3%, MedGemma-27B 82.9%, Llama-3.1-8B 76.3%. The base model before fine-tuning reached 90.5% only with one request per drug; Chaperone-SA-1.0 needs one per document. Test labels were made blind by Claude (Anthropic's model), and a second independent pass agreed on 96% of a sample. No test document was seen in training.
TRAINING Empirisch Tech fine-tuned Chaperone-SA-1.0 from an open-weight base model with low-rank adaptation (LoRA), on examples drawn from a curated corpus of more than 49,000 pharma and medical documents. It runs in FP8 on vLLM 0.29.
SPEED
- ml.g6e.2xlarge (1 NVIDIA L40S, 48 GB): 0.6-1.5 seconds per sample document, measured end to end over the internet from Europe to us-east-1 (4 requests).
- One NVIDIA RTX PRO 6000 Blackwell Max-Q GPU (96 GB), outside SageMaker: about 6-7 documents per second with 32 requests in parallel (the 475 test documents in about 70 seconds).
ON SAGEMAKER AI
- Real-time endpoints and batch transform on one GPU: ml.g6e.2xlarge (recommended) up to ml.g6e.16xlarge, and for real-time also ml.g7e.2xlarge to ml.g7e.8xlarge and ml.p5.4xlarge.
- OpenAI-compatible chat completions with JSON-schema output. Our sample notebook and open-source Python client build each request the way the model was trained.
- Runs network-isolated in your AWS account: your texts and results never reach us, and its logs never contain them.
LIMITATIONS English text; training covered documents of up to about 4,000 tokens. Up to 8 targets per request. Some figures rest on few items: 45 negatives, and 21-29 labels each for trial records, press and social posts.
For research and commercial-intelligence use. Not a medical device, not for diagnosis or treatment decisions, and not a replacement for pharmacovigilance.
Sources: Vals.ai MedQA leaderboard (April 2026); Nori et al. 2023 and 2024; Wang et al. 2025; LoRA: Hu et al. 2021; MedQA: Jin et al. 2020.
Highlights
- Medical knowledge next to frontier models: 94.2% on MedQA (US medical licensing exam, thinking mode on), about 2 points below the top of the Vals.ai leaderboard (o1, 96.5%, April 2026).
- 91% accurate drug-level sentiment on 473 held-out test labels, 8 to 15 points ahead of Qwen3.8-27B, Qwen3-32B, MedGemma-27B and Llama-3.1-8B. Its base model reaches 90.5% only with one request per drug.
- Private by design: runs network-isolated on one GPU in your AWS account (ml.g6e.2xlarge recommended), and returns a valid label for every requested drug: 0 failures in 1,500 test answers.
Details
Sold by
Delivery method
Latest version
Deployed on AWS
New
Introducing multi-product solutions
You can now purchase comprehensive solutions tailored to use cases and industries.
Features and programs
Financing for AWS Marketplace purchases
AWS Marketplace now accepts line of credit payments through the PNC Vendor Finance program. This program is available to select AWS customers in the US, excluding NV, NC, ND, TN, & VT.
Pricing
Free trial
Try this product free for 5 days according to the free trial terms set by the vendor.
Pricing is based on actual usage, with charges varying according to how much you consume. Subscriptions have no end date and may be canceled any time.
Additional AWS infrastructure costs may apply. Use the AWS Pricing Calculator to estimate your infrastructure costs.
Dimension | Description | Cost |
|---|---|---|
ml.g6e.2xlarge Inference (Batch) Recommended | Model inference on the ml.g6e.2xlarge instance type, batch mode | $2.00/host/hour |
ml.g6e.4xlarge Inference (Batch) | Model inference on the ml.g6e.4xlarge instance type, batch mode | $2.00/host/hour |
ml.g6e.8xlarge Inference (Batch) | Model inference on the ml.g6e.8xlarge instance type, batch mode | $2.00/host/hour |
ml.g6e.16xlarge Inference (Batch) | Model inference on the ml.g6e.16xlarge instance type, batch mode | $2.00/host/hour |
inference.count.m.i.c Inference Pricing | inference.count.m.i.c Inference Pricing | $2.00/request |
Dimensions summary
This model package bills by usage. Four dimensions charge by the hour for batch inference, each tied to a different GPU instance size (2xlarge, 4xlarge, 8xlarge, and 16xlarge). You pick the instance that fits your workload, and cost scales with the size you run and the hours you use. A fifth dimension charges per inference request. You run the model inside your own AWS account on Amazon SageMaker AI, so compute scales with how much text you classify.
Top-of-mind questions for buyers
Each hourly dimension maps to one GPU instance size: 2xlarge, 4xlarge, 8xlarge, or 16xlarge. Larger instances give more GPU and memory for heavier batch jobs. You pick one instance, and cost scales with its size and the hours it runs in your AWS account.
The hourly dimensions meter running host-hours. You accrue software charges only while the instance is active and processing. Stopped instances do not run up software charges. Underlying AWS infrastructure fees, such as storage, may still apply separately through AWS.
The four instance dimensions bill by host-hours for batch mode. The fifth dimension bills per inference request. You use the mode that fits your workload. Batch suits large text volumes processed together; request-based suits classifying individual texts. Charges reflect the mode you run.
empirischtech.at
Helpful?
Vendor refund policy
The software fee is charged per hour of use and isn't refundable once incurred. If you think you were charged in error, write to contact@empirischtech.at within 30 days of the charge; we'll review it and, where justified, refund it through AWS Marketplace.
Legal
Vendor terms and conditions
Upon subscribing to this product, you must acknowledge and agree to the terms and conditions outlined in the vendor's End User License Agreement (EULA) .
Content disclaimer
Vendors are responsible for their product descriptions and other product content. AWS does not warrant that vendors' product descriptions or other product content are accurate, complete, reliable, current, or error-free.
Delivery details
Amazon SageMaker model
An Amazon SageMaker model package is a pre-trained machine learning model ready to use without additional training. Use the model package to create a model on Amazon SageMaker for real-time inference or batch processing. Amazon SageMaker is a fully managed platform for building, training, and deploying machine learning models at scale.
Deploy the model on Amazon SageMaker AI using the following options:
Version release notes
First release of Chaperone-SA-1.0 on Amazon SageMaker AI: real-time endpoints on ml.g6e (NVIDIA L40S), ml.g7e and ml.p5.4xlarge, and batch transform on ml.g6e. OpenAI-compatible chat completions with JSON-schema output, served by vLLM 0.29 with FP8 weights. The container's log holds only readiness, request status codes, throughput and error messages, never your texts.
Additional details
Inputs
- Summary
One OpenAI chat-completions request as JSON (content type application/json): the system prompt, word for word; a user message with the source type, the text between <<< and >>>, and 1 to 8 numbered targets with their aliases; temperature 0; thinking off; and a JSON schema with one Positive, Neutral or Negative answer per target. The Python client in the GitHub repository builds it, and samples/request.json shows a complete one.
Batch transform takes a JSON Lines file with one such request per line. Set SplitType to Line, BatchStrategy to SingleRecord and ContentType to application/json.
- Limitations for input type
- English text. 1 to 8 targets per request: split longer lists into several requests. On ml.g6e (48 GB GPU) a request with its answer may hold 8,192 tokens, and on ml.g7e and ml.p5.4xlarge 32,768. Training covered documents of up to about 4,000 tokens, so split longer documents into passages. Keep the system prompt word for word: changing it lowers accuracy. Batch transform sends one request per line (SingleRecord), of up to 6 MB.
- Input MIME type
- application/json
https://github.com/empirischtechinfo/chaperone-med-sagemaker/blob/main/samples/request.json
https://github.com/empirischtechinfo/chaperone-med-sagemaker/blob/main/samples/input.jsonl
Input data descriptions
The following table describes supported input data fields for real-time inference and batch transform.
Field name | Description | Constraints | Required |
|---|---|---|---|
messages | The system prompt, then the user message with the source type, the text and the numbered targets | System message: SYSTEM_PROMPT from the repository's client, word for word, optionally followed by your instructions (up to 2,000 characters) and INSTRUCTIONS_GUARD. User message: the format in samples/request.json. | Yes |
response_format | JSON schema of the answer: one property per target, each Positive, Neutral or Negative | {"type": "json_schema", "json_schema": {"name": "chaperone_sentiment", "schema": ...}} with every target required | Yes |
temperature | Sampling temperature | Use 0 | Yes |
max_tokens | Room for the answer | 32 plus, per target, 16 + half the label's length; at most 512 | No |
top_p | Nucleus sampling | Use 1 | No |
Support
Vendor support
AWS infrastructure support
AWS Support is a one-on-one, fast-response support channel that is staffed 24x7x365 with experienced and technical support engineers. The service helps customers of all sizes and technical abilities to successfully utilize the products and features provided by Amazon Web Services.
Similar products
The Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) data package contains information on medication errors, quality complaints and drug-related adverse events that were submitted to FDA in 2021.
The Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) data package contains information on medication errors, quality complaints and drug-related adverse events that were submitted to FDA in 2017.
The Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) data package contains information on medication errors, quality complaints and drug-related adverse events that were submitted to FDA in 2019.
This data package contains information on Structured Product Labeling (SPL) Terminology for SPL validation procedures and information on performing SPL validations.
The Food and Drug Administration (FDA) Adverse Event Reporting System (FAERS) data package contains information on medication errors, quality complaints and drug-related adverse events that were submitted to FDA in 2018.
Customer reviews
No customer reviews yet
Be the first to review this product . We've partnered with PeerSpot to gather customer feedback. You can share your experience by writing or recording a review, or scheduling a call with a PeerSpot analyst.