MedSafetyBench
Models respond to requests that conflict with medical ethics. A judge rates harmfulness on a five-point scale using a held-out set of 900 requests.
In October 2024, the authors compared Medalpaca, Meditron, ClinicalCamel, Med42, and their Llama 2 counterparts. They report harmfulness by request group and training condition.
- Made by
- Tessa Han and colleagues
- Released
- March 2024
- Kind
- Refusing harmful requests
- Level
- Everyday
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- All public
- Measures
- Propensity
- Risk
- Specialized advice
- Behavior
- Harmful use
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Biology and medicine, Content safety
- Strongest model tested
-
Llama 3.1-70B #177 on the capability index
- Last model added
- July 2024
- Scores refreshed
- November 20th, 2026
llama2-7b-chat tops the leaderboard with 1.01 Llama 2-13B leads with 1.44 among models released by July 2024
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The headline averages the mean harmfulness ratings of the GPT4 and Llama2 medical request groups. Lower is better. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is Llama 2-13B, at 1.44. This view shows only models we track, each at its best.
The newest model with a score here came out in July 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 llama2-7b-chat 1.01
- 2 llama2-70b-chat 1.02
- 3 llama2-13b-chat 1.03
- 4 gpt3.5 (snapshot unspecified) 1.04
- 5 gpt4 (snapshot unspecified) 1.14
- 6 gpt4o (snapshot unspecified) 1.14
- 7 llama3.1-70b-instruct 1.14
- 8 gpt4o-mini (snapshot unspecified) 1.20
- 9 mixtral-8x7b-instruct 1.24
- 10 llama3.1-8b-instruct 1.27
- 11
Llama 2-13B 1.44
- 12
Llama 2-70B 1.59
- 13
Mistral-7B-Instruct-v0.1 1.65
- 14 meditron-70b 1.78
- 15
Llama 2-7B 1.79
- 16 med42-70b 1.89
- 17 pythia-12b 1.91
- 18 clinicalcamel-70b 2.15
- 19 medalpaca-7b 2.24
- 20 medalpaca-13b 2.40
- 21 pythia-6.9b 2.45
- 22
Llama 3.1-70B 2.54
- 23 meditron-7b 2.55
- 24
LLaMA-7B 2.67
- 25
LLaMA-13B 2.73
- 26
Llama 3.1-8B 2.83
- 27
Mistral 7B v0.1 2.87
- 28
Mixtral 8x7B 3.02
- 29 vicuna-7b 3.13
- 30 vicuna-13b 3.17
Source: The MedSafetyBench paper
- 1
Llama 2-13B 1.44
- 2
Llama 2-70B 1.59
- 3
Mistral-7B-Instruct-v0.1 1.65
- 4
Llama 2-7B 1.79
- 5
Llama 3.1-70B 2.54
- 6
LLaMA-7B 2.67
- 7
LLaMA-13B 2.73
- 8
Llama 3.1-8B 2.83
- 9
Mistral 7B v0.1 2.87
- 10
Mixtral 8x7B 3.02
Source: The MedSafetyBench paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | llama2-7b-chat | 1.01 | Official leaderboard | ||
| 2 | llama2-70b-chat | 1.02 | Official leaderboard | ||
| 3 | llama2-13b-chat | 1.03 | Official leaderboard | ||
| 4 | gpt3.5 (snapshot unspecified) | 1.04 | Official leaderboard | ||
| 5 | gpt4 (snapshot unspecified) | 1.14 | Official leaderboard | ||
| 6 | gpt4o (snapshot unspecified) | 1.14 | Official leaderboard | ||
| 7 | llama3.1-70b-instruct | 1.14 | Official leaderboard | ||
| 8 | gpt4o-mini (snapshot unspecified) | 1.20 | Official leaderboard | ||
| 9 | mixtral-8x7b-instruct | 1.24 | Official leaderboard | ||
| 10 | llama3.1-8b-instruct | 1.27 | Official leaderboard | ||
| 11 | | 1.44 | Jul 2023 | Official leaderboard | |
| 12 | | 1.59 | Jul 2023 | Official leaderboard | |
| 13 | | 1.65 | Sep 2023 | Official leaderboard | |
| 14 | meditron-70b | 1.78 | Official leaderboard | ||
| 15 | | 1.79 | Jul 2023 | Official leaderboard | |
| 16 | med42-70b | 1.89 | Official leaderboard | ||
| 17 | pythia-12b | 1.91 | Official leaderboard | ||
| 18 | clinicalcamel-70b | 2.15 | Official leaderboard | ||
| 19 | medalpaca-7b | 2.24 | Official leaderboard | ||
| 20 | medalpaca-13b | 2.40 | Official leaderboard | ||
| 21 | pythia-6.9b | 2.45 | Official leaderboard | ||
| 22 | | 2.54 | Jul 2024 | Official leaderboard | |
| 23 | meditron-7b | 2.55 | Official leaderboard | ||
| 24 | | 2.67 | Feb 2023 | Official leaderboard | |
| 25 | | 2.73 | Feb 2023 | Official leaderboard | |
| 26 | | 2.83 | Jul 2024 | Official leaderboard | |
| 27 | | 2.87 | Sep 2023 | Official leaderboard | |
| 28 | | 3.02 | Dec 2023 | Official leaderboard | |
| 29 | vicuna-7b | 3.13 | Official leaderboard | ||
| 30 | vicuna-13b | 3.17 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 1.44 | Jul 2023 | Official leaderboard | |
| 2 | | 1.59 | Jul 2023 | Official leaderboard | |
| 3 | | 1.65 | Sep 2023 | Official leaderboard | |
| 4 | | 1.79 | Jul 2023 | Official leaderboard | |
| 5 | | 2.54 | Jul 2024 | Official leaderboard | |
| 6 | | 2.67 | Feb 2023 | Official leaderboard | |
| 7 | | 2.73 | Feb 2023 | Official leaderboard | |
| 8 | | 2.83 | Jul 2024 | Official leaderboard | |
| 9 | | 2.87 | Sep 2023 | Official leaderboard | |
| 10 | | 3.02 | Dec 2023 | Official leaderboard |
Rankings by part
We read 2 parts from the benchmark’s paper: MedSafety-Eval-GPT4 and MedSafety-Eval-Llama2. Each ranks models on one part of the test. Llama 2-13B leads MedSafety-Eval-GPT4 and Mistral-7B-Instruct-v0.1 leads MedSafety-Eval-Llama2. Arrows show how far a model moved from its overall rank among the models we track.
MedSafety-Eval-GPT4
10 models we track have a score here.
- 1
Llama 2-13B 1.51
- 2
Llama 2-70B 1.65
- 3
Llama 2-7B ↑1 1.86
- 4
Mistral-7B-Instruct-v0.1 ↓1 1.95
- 5
LLaMA-7B ↑1 2.91
- 6
Llama 3.1-70B ↓1 3.02
- 7
Mistral 7B v0.1 ↑2 3.07
- 8
LLaMA-13B ↓1 3.13
- 9
Llama 3.1-8B ↓1 3.18
- 10
Mixtral 8x7B 3.30
MedSafety-Eval-Llama2
10 models we track have a score here.
- 1
Mistral-7B-Instruct-v0.1 ↑2 1.34
- 2
Llama 2-13B ↓1 1.36
- 3
Llama 2-70B ↓1 1.54
- 4
Llama 2-7B 1.73
- 5
Llama 3.1-70B 2.06
- 6
LLaMA-13B ↑1 2.34
- 7
LLaMA-7B ↓1 2.43
- 8
Llama 3.1-8B 2.47
- 9
Mistral 7B v0.1 2.67
- 10
Mixtral 8x7B 2.73
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The MedSafetyBench paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official MedSafetyBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6%
- Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts.
Gemini 4 Carbon 0%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 2 years ago, in July 2024, for Llama 3.1-70B. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (30 entries).
- Pass Models scored 10 models we track, plus 20 other entries on its leaderboard.
- Warning Newest score The newest came 2 years ago, in July 2024, for Llama 3.1-70B.
- Note Frontier models None of the top 10 has a score. The strongest tested is Llama 3.1-70B (#177).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on proceedings.neurips.cc: 10 rows for models we track and 20 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.