Create bio_evals_and_llms.txt

bio_evals_and_llms.txt

@@ -0,0 +1,119 @@
+Biology Benchmarks & Models for Generative LLMs
+=================================================
+
+## Benchmarks for Evaluating Base/Generative LLMs on Biology
+
+### Knowledge-Focused (Multiple Choice)
+
+1. MMLU Biology Subsets
+   - Includes: college biology, high school biology, genetics, anatomy, etc.
+   - Format: multiple-choice
+   - Note: somewhat saturated; GPT-3.5 Turbo was already >70%
+   - Good as an easy baseline
+
+2. GPQA Biology (Molecular Biology & Genetics)
+   - 78 graduate-level, Google-proof questions
+   - Very hard: human PhD experts score ~66.7%, non-expert PhDs ~43.2%
+   - One of the best current discriminators for frontier models
+   - https://arxiv.org/pdf/2311.12022
+
+3. WMDP-Bio
+   - 1,273 questions covering biosecurity-adjacent bio knowledge
+   - Topics: reverse genetics, viral vectors, enhanced pandemic pathogens
+   - Human PhD experts scored ~60.5%
+   - https://www.rand.org/pubs/research_reports/RRA3797-1.html
+
+4. SciEval (Biology subset)
+   - ~18,000 questions across chemistry, physics, and biology
+   - Organized by Bloom's taxonomy: basic knowledge -> research ability
+   - Includes a dynamic subset to prevent data leakage
+   - https://github.com/OpenDFM/SciEval
+
+5. SciKnowEval
+   - Subsets: biology, chemistry, materials, physics
+   - Multiple question types: relation-extraction, MCQ, true/false, generation
+   - https://arxiv.org/abs/2406.09098
+
+6. Humanity's Last Exam (HLE)
+   - 2,500 expert-level questions, ~11% biology/medicine
+   - Extremely hard, designed to challenge frontier models
+
+### QA / Comprehension
+
+7. PubMedQA
+   - Yes/no/maybe answers to research questions from PubMed abstracts
+   - Tests literature comprehension of biomedical text
+
+8. BioASQ
+   - Annual biomedical QA challenge
+   - Factoid, list, and yes/no questions from biomedical experts
+
+9. MedQA
+   - Medical board exam multiple-choice questions
+
+### Practical / Applied Biology
+
+10. LAB-Bench (Language Agent Biology Benchmark)
+    - 2,400+ MCQs testing practical biology research capabilities
+    - SeqQA subtask: 15 subtasks on DNA/protein sequence comprehension
+    - CloningScenarios: complex multi-step molecular cloning problems
+    - https://arxiv.org/abs/2407.10362
+
+11. BixBench
+    - 50+ real-world computational biology scenarios, ~300 open-answer questions
+    - Tests multi-step analytical trajectories on real biological datasets
+    - Very hard: frontier models only achieve ~17% accuracy
+    - https://arxiv.org/abs/2503.00096
+
+### Recommended Evaluation Suite (from easy to hard)
+- MMLU-bio (easy baseline)
+- GPQA-bio (hard discriminator)
+- PubMedQA (comprehension)
+- LAB-Bench (practical DNA/cloning tasks)
+- HLE biology subset (frontier-hard)
+- BixBench (agentic, very hard)
+
+
+## Generative LLMs Fine-Tuned on Biology / Biomedicine
+
+1. BioMistral (Mistral-7B based)
+   - Further pre-trained on PubMed Central
+   - Open-source, designed for biomedical text generation
+   - https://huggingface.co/BioMistral/BioMistral-7B
+
+2. BioMed-LLaMa-3 (Llama-3-8B based)
+   - Instruction-tuned on ~54K biomedical examples
+   - https://link.springer.com/chapter/10.1007/978-981-96-0695-5_32
+
+3. BioQwen (Qwen 0.5B/1.5B/1.8B based)
+   - Bilingual (Chinese/English) biomedical, two-stage fine-tuning
+   - Competitive with larger models on biomedical generation tasks
+
+4. ChatMultiOmics (Llama-based)
+   - Instruction-tuned on DNA/RNA/protein sequence tasks
+   - Uses the Biology-Instructions dataset (first large-scale multi-omics
+     instruction-tuning dataset)
+   - https://arxiv.org/abs/2412.19191
+
+5. MedBioLM
+   - Fine-tuned + RAG approach
+   - 88% on MedQA, 78.9% on PubMedQA, 96% on BioASQ
+   - https://arxiv.org/abs/2502.03004
+
+Note: For raw DNA sequence understanding (not natural language about biology),
+specialized genomic foundation models exist in a separate category:
+- Evo2 (40B params, trained on 9.3T DNA base pairs, 1M context window)
+- DNABERT-2
+These operate on nucleotide sequences directly, not natural language.
+
+
+## Key References
+
+- LLMs Outperform Experts on Challenging Biology Benchmarks (MIT, 2025)
+  https://arxiv.org/abs/2505.06108
+- Trendlines in AIxBio Evals
+  https://www.lennijusten.com/blog/biology-benchmarks/
+- LLM Benchmarks in Life Sciences Overview
+  https://intuitionlabs.ai/articles/large-language-model-benchmarks-life-sciences-overview
+- Bridging AI and Biological Sciences (Briefings in Bioinformatics, 2025)
+  https://academic.oup.com/bib/article/26/4/bbaf357/8212018