<think> and </think> are atomic non-special added tokens. This keeps the markers visible to normal decoders and reasoning-aware chat interfaces.

Expected assistant response:

<think>
Short, relevant reasoning.
</think>
Final answer

Training

Context length: 4096
Train rows: 28,786
Validation rows: 1,000
Learning rate: 1.5e-05
Epochs: 2.0
Reasoning loss weight: 1.0
Final-answer loss weight: 4.0
Seed: 3407

Ma7ee7/SmolLM2-135M-Reasoning-Recovery-30K-4096

SmolLM2-135M Reasoning Recovery 30K — 4096

Reasoning-token format

Training

Ma7ee7/
SmolLM2-135M-Reasoning-Recovery-30K-4096