EAGLE3-Apertus-8B-Instruct-2509-Data

Training dataset for the thomaskiefer/EAGLE3-Apertus-8B-Instruct-2509 speculative decoding draft model.

Dataset Description

This dataset contains ~375k multi-turn conversations used to train an Eagle3 draft model for swiss-ai/Apertus-8B-Instruct-2509.

Data Sources

The prompts are sourced from:

• UltraChat - Large-scale multi-turn dialogue dataset
• ShareGPT - Real user conversations
• OpenThoughts-114k-math - Mathematical reasoning data

Regeneration Process

The responses were regenerated using Apertus-8B-Instruct-2509 rather than using the original responses. This ensures the draft model learns to predict tokens from the target model's own output distribution, which is critical for effective speculative decoding.

Dataset Format

JSONL format with the following structure:

{
  "id": "unique_sha256_hash",
  "conversations": [
    {"role": "user", "content": "User message..."},
    {"role": "assistant", "content": "Regenerated assistant response..."},
    {"role": "user", "content": "Follow-up question..."},
    {"role": "assistant", "content": "Regenerated follow-up response..."}
  ],
  "status": "success"
}

Fields

Field	Type	Description
id	string	SHA-256 hash identifier for the conversation
conversations	array	List of conversation turns with role and content
status	string	Processing status (success indicates valid sample)

Dataset Statistics

• Format: JSONL
• Samples: 375,573
• Languages: Primarily English

Usage

Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("thomaskiefer/EAGLE3-Apertus-8B-Instruct-2509-Data")

Load Directly

import json

conversations = []
with open("merged_train_regen.jsonl", "r") as f:
    for line in f:
        conversations.append(json.loads(line))

Train with SpecForge

NUM_GPUS=4
TP_SIZE=1

torchrun \
    --standalone \
    --nproc_per_node $NUM_GPUS \
    scripts/train_eagle3.py \
    --target-model-path swiss-ai/Apertus-8B-Instruct-2509 \
    --draft-model-config /path/to/configs/apertus-8b-eagle3.json \
    --train-data-path /path/to/merged_train_regen.jsonl \
    --output-dir /path/to/outputs/apertus-8b-eagle3 \
    --num-epochs 10 \
    --batch-size 1 \
    --tp-size $TP_SIZE \
    --learning-rate 1e-4 \
    --max-length 4096 \
    --chat-template apertus \
    --cache-dir /path/to/cache \
    --target-model-backend sglang

Related Resources

• EAGLE3-Apertus-8B-Instruct-2509 - Trained draft model
• SpecForge - Training framework
• Eagle3 Paper
• Apertus-8B-Instruct-2509 - Target model

License

Apache 2.0

Citation

@article{li2025eagle3,
  title={Eagle 3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test},
  author={Li, Yuhui and Wei, Fangyun and Zhang, Chao and Zhang, Hongyang},
  journal={arXiv preprint arXiv:2503.01840},
  year={2025}
}