Add dataset card

README.md

@@ -0,0 +1,112 @@
+---
+language:
+- en
+license: apache-2.0
+size_categories:
+- 100K<n<1M
+task_categories:
+- text-generation
+- question-answering
+pretty_name: Science QA SFT (100K)
+tags:
+- science
+- stem
+- physics
+- chemistry
+- biology
+- astronomy
+- earth-science
+- sft
+- supervised-fine-tuning
+- chain-of-thought
+- step-by-step
+- reasoning
+- synthetic
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: science-qa-sft-100k.jsonl
+---
+
+# Science QA SFT (100K)
+
+100,000 science Q&A examples with step-by-step explanations for SFT fine-tuning. Covers physics, chemistry, biology, astronomy, and earth science at beginner through advanced difficulty.
+
+## Motivation
+
+Models trained on general text often give superficially plausible but mechanistically wrong answers to science questions — stating the right conclusion without understanding the underlying reasoning. This dataset trains models to explain *why* an answer is correct using first principles and step-by-step logic.
+
+## Dataset Description
+
+**100,000 Q&A pairs** across 5 science subjects and 3 difficulty levels:
+
+### Subject Distribution
+
+| Subject | Count | % |
+|---|---|---|
+| Physics | ~31,000 | 31% |
+| Chemistry | ~19,300 | 19% |
+| Biology | ~19,200 | 19% |
+| Astronomy | ~15,300 | 15% |
+| Earth Science | ~15,200 | 15% |
+
+### Difficulty Distribution
+
+| Difficulty | Count | % |
+|---|---|---|
+| Easy | ~34,600 | 35% |
+| Medium | ~46,200 | 46% |
+| Hard | ~19,200 | 19% |
+
+### Topics Covered
+
+**Physics**: Newton's laws, kinematics, elastic/inelastic collisions, thermodynamics, Ohm's law, electromagnetic induction, wave optics, Heisenberg uncertainty principle
+
+**Chemistry**: Atomic structure, ionic vs covalent bonding, Le Chatelier's principle, acid-base chemistry (pH), hydrocarbons
+
+**Biology**: Prokaryotes vs eukaryotes, DNA replication (semi-conservative model), natural selection, cardiovascular circulation, ecological succession
+
+**Astronomy**: Seasons (axial tilt), stellar lifecycle (main sequence to stellar remnants), Big Bang evidence (CMB, redshift, abundance), black holes and event horizons
+
+**Earth Science**: Plate tectonics, greenhouse effect and radiative forcing, thermohaline circulation (ocean currents), lightning formation
+
+## Format
+
+```json
+{
+  "question": "Why does ice float on water?",
+  "explanation": "Most substances are denser in solid form than liquid form because molecules pack more tightly when frozen. Water is unusual: when it freezes, water molecules form a hexagonal crystalline lattice held together by hydrogen bonds. This lattice structure spaces molecules farther apart than in liquid water.\n\nStep 1 — Density comparison: Ice has a density of ~0.917 g/cm³; liquid water is ~1.0 g/cm³. Since ice is less dense, it floats.\n\nStep 2 — Why the lattice forms: Each water molecule can form 4 hydrogen bonds (2 as donor, 2 as acceptor). In the ice lattice, all 4 bonds are satisfied in a rigid, open hexagonal ring — creating more space between molecules than in the disordered liquid state.\n\nStep 3 — Ecological significance: This property is crucial for aquatic life. Ice insulates liquid water below, preventing lakes from freezing solid in winter. If ice sank, bodies of water would freeze from the bottom up, eliminating most aquatic ecosystems in cold climates.",
+  "answer": "Ice floats because its crystalline lattice structure (formed by hydrogen bonding) makes it less dense (~0.917 g/cm³) than liquid water (~1.0 g/cm³). This anomalous property is essential for aquatic ecosystems in cold climates.",
+  "subject": "chemistry",
+  "topic": "hydrogen bonding",
+  "difficulty": "medium",
+  "id": "sci_abc123"
+}
+```
+
+## Key Training Signals
+
+1. **Mechanistic reasoning** — answers trace causal chains, not just state facts
+2. **Quantitative grounding** — specific values (densities, forces, temperatures) anchor explanations
+3. **Analogies and context** — connect abstract concepts to everyday phenomena
+4. **Multi-step structure** — explanations break complex reasoning into labeled steps
+5. **Difficulty calibration** — easy answers are direct; hard answers address misconceptions and edge cases
+
+## Difficulty Definitions
+
+- **Easy**: Recall + single-step reasoning; appropriate for middle/high school level
+- **Medium**: Multi-step reasoning with quantitative elements; appropriate for introductory college level
+- **Hard**: Addresses common misconceptions explicitly; requires integrating 2+ concepts; appropriate for upper-division undergraduate level
+
+## Use Cases
+
+- SFT fine-tuning for STEM question answering
+- Training models to produce step-by-step explanations vs. bare answers
+- Science tutoring applications requiring pedagogically sound explanations
+- Evaluation of scientific reasoning capabilities
+- Curriculum learning: train on easy → medium → hard progressively
+
+## License
+
+Apache 2.0