| # Genue Matrix 8B | |
| ## Model Overview | |
| Genue Matrix 8B is a specialized reasoning engine designed for algebraic word problems, symbolic logic, and mathematical proofs. It leverages a weighted multi-source training strategy to maximize its logical deduction capabilities in a compact 8B parameter package. | |
| ## The Breakthrough | |
| This model achieved a record-low training loss of 0.35 on the MathInstruct dataset, a highly competitive benchmark for reasoning models. By utilizing a Cosine Soft-Landing scheduler and TIGER-DNA data weighting, Genue Matrix maintains coherence in complex multi-step derivations where standard 8B models typically experience logic collapse. | |
| ## Training Details | |
| - **Base Model**: Llama 3 8B (via Genue Matrix Prime) | |
| - **Dataset**: MathInstruct (2,000 weighted samples of GSM8K, MATH, AQuA, and ProofWiki) | |
| - **Final Training Loss**: 0.35 | |
| - **Scheduler**: Cosine Annealing (for superior convergence) | |
| ## Performance & Tests | |
| - **Algebraic Reasoning**: High (Verified via MathInstruct) | |
| - **Word Problem Logic**: High (Sub-0.40 loss convergence) | |
| - **Symbolic Consistency**: High | |
| ## Usage | |
| The model follows the Alpaca instruction format: | |
| ``` | |
| ### Instruction: | |
| [Your algebraic or logic problem] | |
| ### Response: | |
| ``` | |