Announcing ALR Aerospace 0.2: Deterministic Orbital Mechanics at Scale
Today, we are releasing ALR Aerospace 0.2—a complete rebuild of our aerospace reasoning dataset. This release marks a fundamental shift in how synthetic data is generated for the aerospace industry.
Most LLMs struggle with orbital mechanics, propulsion, and atmospheric physics because they rely on probabilistic text generation. They guess the math based on patterns in their training data. ALR Aerospace 0.2 does not guess. It computes.
Here is a comprehensive breakdown of what is new, what was fixed, and how to integrate the 0.2 release into your training pipeline.
1. The Deduplication Problem (v0.1 vs v0.2)
The alr-aerospace-platform (v0.1) contained over 13,400 rows. However, during our independent audit, we discovered that nearly 60% of these rows were duplicate fuzzer variations. We had 2,614 unique problem IDs repeating exactly four times each.
While these duplicates stayed safely within their train/test splits (zero leakage), they artificially inflated the dataset size and wasted compute during fine-tuning. Training on the same problem four times in one epoch leads to overfitting and wasted tokens.
The Fix: alr-aerospace-platform-0.2 is 100% deduplicated. Every single row is completely unique. The dataset has been strictly deduplicated by sample_id, instance_key, and provenance_hash. When you stream Aerospace 0.2, you are training on 4,600 entirely distinct, non-repeating problem instances.
2. Three New High-Fidelity Problem Families
Aerospace 0.2 expands the platform to 31 unique problem templates, introducing three highly requested, mathematically rigorous domains:
isa1976_full (Atmosphere): The complete US Standard Atmosphere 1976 model, spanning all 7 layers from 0 km to 84.852 km. For any given altitude, the engine calculates density, pressure, temperature, and the speed of sound, accounting for the exact lapse rates and molecular weights per layer.
gravity_assist_flyby (Interplanetary Design): Patched-conic planetary gravity assists. The engine calculates the heliocentric velocity change and turn angle based on planetary mass, flyby altitude, and the incoming v-infinity vector. This is critical for teaching models how to design multi-planet trajectories.
geo_transfer_budget (Mission Delta-v): LEO-to-GEO transfers utilizing Hohmann transfers with plane changes folded directly into the apogee burn. It includes the exact delta-v budget, teaching models how to calculate the massive cost of changing orbital inclination.
3. Certified Enclosures and Triple-Integrator Agreement
The core philosophy of Alrithm is that data must be mathematically irrefutable. Every row in Aerospace 0.2 falls into one of two strict fidelity tiers:
Tier 1: Certified Interval Enclosures. Closed-form orbital problems (e.g., Hohmann transfers, Tsiolkovsky rocket equation) are verified using mpmath.iv 60-digit interval arithmetic. Every row stores a rigorous numerical enclosure in the headline_answer_exact field, eliminating floating-point uncertainty entirely. The relative half-width is approximately 10^-60.
Tier 2: Converged Numerical Trajectories. Trajectory ODE problems (e.g., ascent gravity turns, J2 propagation) are cross-checked across three independent integrators: fixed-step RK4, explicit 8th-order Dormand-Prince (DOP853), and implicit Radau. The rows are only published if all three integrators agree to within < 10^-11 relative error. This guarantees the solution is not an artifact of a single integrator's step-size or stiffness handling.
Furthermore, every row explicitly documents its physical assumptions (e.g., GM = 3.986004418e14, Re = 6378137) and lists its neglected_effects (e.g., Earth rotation, Mach-dependent drag, winds).
4. How to Stream the New Dataset
You can access Aerospace 0.2 immediately using your existing API key.
To stream the raw, fully verified JSONL format (which includes the verified_solution_code, the Merkle proofs, and the step-by-step reasoning), use this cURL command:
curl -N "https://api.alrithm.com/v1/models/alr-aerospace-platform-0.2/data?format=raw"
-H "Authorization: Bearer sk_live_your_key_here" \
aerospace_training.jsonl
If you are piping the data directly into an OpenAI fine-tuning job, we recommend using the pre-formatted serializer by changing ?format=raw to ?format=openai.
5. What is Next?
We will be publishing a full benchmark report comparing models fine-tuned on Aerospace 0.2 against models trained on raw aerospace literature. We expect to see a significant reduction in hallucination rates for orbital trajectory calculations.
As always, if you discover any edge cases or physical anomalies in the dataset, report them via our feedback channel. The engine is only as good as the physics it verifies.
Associated Data Model
This article references the ALR Aerospace Platform 0.2. You can stream these verified reasoning traces immediately using our API.
[ FETCH DATA VIA API ]