---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - distillation
  - olmo
  - instruction-tuning
---

# Dolci Instruct RL Completions

Instruction-following completions sampled from OLMo-3-7B-Instruct with teacher logits for knowledge distillation.

## Description

This dataset contains instruction-completion pairs with pre-computed teacher logits from OLMo-3-7B-Instruct. Designed for training smaller student models via KL-divergence distillation.

## Generation

Completions were sampled from **allenai/OLMo-3-7B-Instruct** on instruction prompts. For each token position, we store the top-128 token indices and their corresponding logits.

## Format

Each example contains:
- `input_ids_prompt`: tokenized instruction/prompt
- `input_ids_completion`: tokenized model completion
- `teacher_indices`: [seq_len, 128] top-128 token indices per position
- `teacher_logits`: [seq_len, 128] corresponding logits

## Usage



## Pre-packed Version

For ready-to-train packed batches, see [hbfreed/dolci-distill-packed](https://huggingface.co/datasets/hbfreed/dolci-distill-packed).