datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
eagle3-hidden-states-gemma4llama4-sglang-eagle3qwen3_8b_eagle3-parquet
qwen3_8b_eagle3 (Parquet)
Sharded Parquet conversion of Tengyunw/qwen3_8b_eagle3.
Original distribution is a single ~13 GB JSON file; this repo splits it into
61 Parquet shards of ~10,000 rows each for streaming-friendly access
via the datasets library.
Schema
id: string
conversations: list<struct<from: string, value: string>> (ShareGPT format)
Stats
Rows: 607,865
Shards: 61 (data/train-NNNNN-of-00061.parquet)
Compression: zstd
Usage
from… See the full description on the dataset page: https://huggingface.co/datasets/shadowpa0327/qwen3_8b_eagle3-parquet.eagle3-speculative-decoding-energy-sweep
EAGLE3 Speculative Decoding Energy Sweep
Per-config energy/throughput/latency measurements for EAGLE3 speculative decoding
(speculative_num_steps, speculative_eagle_topk, speculative_num_draft_tokens)
served with sglang, across batch sizes. Collected for an RL project that learns to
pick speculative-decoding parameters to hold GPU energy utilization in a target band.
Model: unsloth/Llama-3.2-1B-Instruct + rescommons/SpecForge-EAGLE3-Llama-3.2-1B-Instruct draft head.
Hardware:… See the full description on the dataset page: https://huggingface.co/datasets/Pradheep1647/eagle3-speculative-decoding-energy-sweep.angelslim-smolvlm-eagle3-artifactsEAGLE3-Apertus-8B-Instruct-2509-Data
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… See the full description on the dataset page: https://huggingface.co/datasets/thomaskiefer/EAGLE3-Apertus-8B-Instruct-2509-Data.MoS-Qwen3-8B-EAGLE3-responses
MoS — Qwen3-8B EAGLE3 Training Responses
Target-model responses for training EAGLE3 speculative-decoding draft models against
Qwen/Qwen3-8B. Built for the MoS (Mixture of
Speculators) project — a routed multi-MLP draft — and equally usable for any single-draft
EAGLE3 / SpecForge training run on Qwen3-8B.
599,087 complete assistant responses (with thinking traces) over five domains, generated
by Qwen3-8B itself so the draft learns to mimic the target's own distribution.… See the full description on the dataset page: https://huggingface.co/datasets/ryan-0608/MoS-Qwen3-8B-EAGLE3-responses.qwen36-eagle3-stagebeagle360_test
EAGLE-360 Test Set
Project page: EAGLE-360
Paper: arXiv:2607.02479
EAGLE-360 is a benchmark for embodied active global-to-local exploration in 360-degree panoramic scenes. Given a panoramic image and a target-object query, the model is asked to predict the object's angular position as azimuth and elevation in degrees.
This release contains the public test split only. It includes panoramic images and a annotation file with ground-truth metadata.
Files… See the full description on the dataset page: https://huggingface.co/datasets/Sansjudge/eagle360_test.eagle3-sarvam-30b-training-data
Eagle3 Sarvam-30B Training Data
Training data used to build the Eagle3 draft model for Sarvam-30B.
Dataset Description
This dataset contains 90,000 prompt-response pairs used to train an Eagle3 speculative decoding draft model for the Sarvam-30B language model.
Each sample consists of a prompt and its corresponding response generated by the Sarvam-30B base model. During training, the model also consumes hidden state features extracted from auxiliary layers of the base… See the full description on the dataset page: https://huggingface.co/datasets/sulabhkatiyar/eagle3-sarvam-30b-training-data.atc-parser-eagle3-data-v1qwen3_8b_eagle3This is the data we regenerated using Qwen3-8B on UltraChat_200K, totaling approximately 600K dialogues and 1B tokens.
The data format is:
json
{"id": xxx, "conversations": [{"from": xx, "value": xxx}]}
c2_eagle3_train
EAGLE3 Training Data — Nemotron-Cascade-2-30B-A3B
Training corpus for an EAGLE3 draft head trained against
nvidia/Nemotron-Cascade-2-30B-A3B
with SpecForge (online training,
HF backend).
All files are JSONL in SpecForge's online-training schema:
{
"id": "...",
"conversations": [{"role": "system|user|assistant|tool", "content": "...", "tool_calls": [...]}],
"tools": [{"type": "function", "function": {...}}]
}
The Nemotron tokenizer's apply_chat_template is used at training… See the full description on the dataset page: https://huggingface.co/datasets/chankhavu/c2_eagle3_train.qwen2.5-0.5b-eagle3-datavllm-0171-eagle3-patched-wheeleagle3-gemma4surrogate-2-eagle3-traceseagle3_training_data