Update model card: v3 training — 4 passes, 6 new engineering domains, 23,850 examples

README.md

@@ -1,13 +1,101 @@
 ---
+language:
+- en
 license: gemma
+base_model: google/gemma-3-12b-it
 library_name: mlx
 pipeline_tag: text-generation
-extra_gated_heading: Access Gemma on Hugging Face
-extra_gated_prompt: To access Gemma on Hugging Face, you’re required to review and
-  agree to Google’s usage license. To do this, please ensure you’re logged in to Hugging
-  Face and click below. Requests are processed immediately.
-extra_gated_button_content: Acknowledge license
-base_model: mlx-community/gemma-3-12b-it-4bit
 tags:
 - mlx
+- aerospace
+- engineering
+- thermodynamics
+- mathematics
+- finance
+- coding
+- signals
+- statics
+- dynamics
+- mechanics-of-materials
+- controls
+- manufacturing
+- lora
+- fine-tuned
+- chain-of-thought
 ---
+
+# gemma3-12b-engineering
+
+A fine-tuned version of [Gemma 3 12B IT](https://huggingface.co/google/gemma-3-12b-it) specialized for **aerospace engineering, thermodynamics, advanced mathematics, coding, finance, and 6 additional engineering disciplines**.
+
+## Model Details
+
+- **Base model:** google/gemma-3-12b-it (4-bit quantized via MLX)
+- **Fine-tuning method:** QLoRA (MLX/LoRA) — 4 sequential training passes
+- **Format:** MLX 4-bit quantized safetensors (~6.7 GB)
+- **Hardware:** Apple MacBook Air M4 16GB
+
+## Training Summary
+
+| Pass | Dataset | Examples | Best Val Loss |
+|------|---------|----------|---------------|
+| v2 domain | MetaMathQA, Open-Platypus, OpenHermes STEM, ArXiv QA, SciQ, WikiText, WikiQA, CAMEL Physics/Math, CodeAlpaca, Finance | ~56K | 0.617 |
+| CoT reasoning | nvidia/OpenMathReasoning, Open-Platypus CoT, MetaMathQA CoT, handcrafted aerospace | ~11K | 0.439 |
+| Precision | Handcrafted aerospace/thermo — correct R=8314/M derivation (never R=287 for custom propellants) | ~60 | 0.620 |
+| v3 comprehensive | 6 new engineering domains + NuminaMath-CoT, Magicoder, Finance-Alpaca, OpenHermes STEM, CodeFeedback | 23,850 | 0.689 |
+
+**LoRA config:** rank=16, alpha=32, lora_layers=4, keys=[q_proj, v_proj], LR=2e-6
+
+## Capabilities
+
+- **Aerospace:** Isentropic flow, normal shocks, Brayton/Rankine cycles, rocket nozzles, Hohmann transfers
+- **Thermodynamics:** Carnot, heat exchangers, entropy, propellant property derivation
+- **Signals & Systems:** Laplace transforms, Bode plots, Z-transforms, Fourier series, stability, RC filters, sampling
+- **Statics:** Equilibrium, beam reactions, trusses, centroids, moments of inertia, friction, frames
+- **Dynamics:** Kinematics, Newton's 2nd law, work-energy, impulse-momentum, rotation, vibrations
+- **Mechanics of Materials:** Axial stress, torsion, bending, Mohr's circle, Euler buckling, thermal stress, deflection
+- **Controls:** Routh-Hurwitz, PID design, state-space, root locus, steady-state error, block diagrams, time specs
+- **Manufacturing:** Turning/milling, Taylor tool life, tolerances, Chvorinov's rule, grinding, machining time
+- **Mathematics:** ODEs, linear algebra, RK4, Newton-Raphson, induction proofs, competition-level math
+- **Coding:** Python, C++, Java, JavaScript, numerical solvers
+- **Finance:** Black-Scholes, NPV, DCF, engineering economics, portfolio theory
+
+## Chain-of-Thought Reasoning
+
+Activate step-by-step reasoning with this system prompt:
+
+```
+You are an expert aerospace engineer. Always reason step by step inside <think> tags before giving your final answer.
+```
+
+## Usage (MLX on Apple Silicon)
+
+```python
+from mlx_lm import load, generate
+
+model, tokenizer = load("vininhosts/gemma3-12b-engineering")
+
+prompt = "A rocket nozzle has Pc=2MPa, Tc=3000K, exit Mach=3, propellant M=20g/mol, gamma=1.3. Find exit pressure."
+response = generate(model, tokenizer, prompt=prompt, max_tokens=1024)
+print(response)
+```
+
+## Key Precision: R = 8314 / M
+
+This model correctly computes the specific gas constant from molar mass as **R = 8314 / M**,
+and never defaults to R = 287 J/(kg·K) (air) when a different propellant molar mass is given.
+This was enforced via a dedicated precision fine-tuning pass with 60+ handcrafted examples.
+
+## Example Domains Covered
+
+- Isentropic nozzle flow and normal shock relations
+- Brayton cycle thermal efficiency and compressor work
+- PID controller tuning and Routh-Hurwitz stability
+- Beam deflection and Mohr's circle stress analysis
+- Z-transform and discrete-time system stability
+- Taylor tool life equation and machining parameters
+- Black-Scholes option pricing and DCF valuation
+
+## License
+
+Derived from Gemma 3 — subject to [Gemma Terms of Use](https://ai.google.dev/gemma/terms).