Distillation of linguistic reasoning

[eduardosanchezg]

We want to get high quality reasoning traces for linguistic reasoning.

Models:

• GPT-OSS 120B
• Qwen 3.5 397B

Benchmarks to distill:

• Lingoly
• IOLBENCH
• LOBSTER

[Sara-Rajaee]

There might be some technical difficulties in mixing reasoning traces distilled from different models. Let's start with Gpt-oss-120b and see how many data points we will end up with.