---
language:
- en
library_name: transformers
tags:
- qwen
- distillation
- ifeval
- retaining-by-doing
---

# Qwen2.5-0.5B Instruct IFEval Mixed KD (alpha=0.5)

This checkpoint was distilled from a half-epoch `Qwen2.5-1.5B-Instruct` teacher trained on `IFEvalSFTDataset`.

Distillation setup:
- student: `Qwen2.5-0.5B-Instruct`
- teacher: half-epoch `Qwen2.5-1.5B-Instruct`
- `num_train_datapoints=4064`
- `num_epochs=1`
- `distill_alpha=0.5`
- `distill_temperature=2.0`
- effective loss mix: `CE 0.5 / KD 0.5`

Observed local IFEval accuracy:
- `0.4137577002`