This model was fine‑tuned with GRPO for only 50 steps using 4 samples per step. The result is exceptionally high accuracy on JEE‑level mathematics problems, though its broader context handling and instruction‑following abilities were diminished. In essence, it has become a compact powerhouse — a “mini‑tank” built for raw mathematical problem‑solving rather than nuanced reasoning.
- Downloads last month
- 6