datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
function-calling-eval-dataset-v0The hf dataset contains 2 evaluation datasets
single_turn - The converstaion length for this evaluation dataset is 2. It consists of a user ask followed by a function call by assistant.
multi_turn - The conversation length is variable here but contains a combination of user messages, assistant function calls, assistant messages & tool responses.
Information about the columns
tools - List of functions/tools with specs in JSON format. This is the list of functions the model has to choose from… See the full description on the dataset page: https://huggingface.co/datasets/fireworks-ai/function-calling-eval-dataset-v0.logiqalogiqa-deepseek-v3bfcl_v3_multi_turn_basejapanese-aerial-fireworks-v2
🎆 NEW: Curated 1,000 Wide Pack (Commercial License)
For commercial AI/ML training, check out the Hanabi AI Dataset v1: Wide Pack — Curated 1,000 — a carefully selected subset with detailed structured annotations:
✅ 1,000 hand-curated 4K images (vs 2,557 raw images here)
✅ Structured AI annotations (composition, mood, color, EXIF, English notes)
✅ Sample PyTorch loader, attribute filter, caption generator
✅ Perpetual Commercial License (Japanese law)
✅ Optimized for Stable… See the full description on the dataset page: https://huggingface.co/datasets/dfhjs2577/japanese-aerial-fireworks-v2.msmarco_rank
Dataset Card for "msmarco_rank"
More Information needed
rllm-tb-v2-debug-fireworks-batchfunction-calling-intent-eval-v1This dataset contains the intent evaluation of fw function calling mode vs GPT-4. The dataset contains both
fw model responses under completion
GPT-4 model responses under previous_completion
GPT-4 acts as a teach and is given the following instructions.
GPT-4 teacher respones are stored under
validation_result
completion_reason/completion_score - GPT-4's reason for giving completion_score to the fw function calling model.
previous_completion_reason/previous_completion_score - GPT-4's… See the full description on the dataset page: https://huggingface.co/datasets/fireworks-ai/function-calling-intent-eval-v1.long-chatfireworks-night
FireWorks Dataset
This dataset contains free-licensed images, downloaded from unsplash. Curated and created by:
DESIGNECOLOGIST
Zoltán Cse
Rene Bernal
kazuend
Leo_Visions
Camille Couvez
Paul Morley
Vernon Raineil Cenzon
Carson Arias
Sebastian Davenport-Handley
Kimson Doan
Nitish Meena
japanese-aerial-fireworks
Japanese Aerial Fireworks Dataset
High-quality aerial fireworks photography dataset for AI training (LoRA / fine-tuning).
Image Details
Format: JPEG
Resolution: up to 1024px (long edge)
Camera: Canon EOS 80D / R10
Time of capture: evening/night (17:00–21:00 JST)
Caption Format
Each .txt file contains a caption starting with the trigger word:
hanabi fireworks, [description of the image]
Usage
Ideal for training LoRA models with Flux, SDXL, or… See the full description on the dataset page: https://huggingface.co/datasets/dfhjs2577/japanese-aerial-fireworks.vision-food-reasoning-datasethanabi-fireworks-state-tracking
FIREWORKS: Hanabi belief-state reconstruction
Strict hidden-belief reconstruction for Hanabi. Each example gives a previous belief state
plus the actions taken since, and asks for the updated per-card possibility sets.
10,185 examples (9,780 unique prompts)
2-5 player games, seeds 101-110 (evaluation seeds 1001-1010 are held out)
Fields: id, meta (num_players, seed, turn, observer, log), prompt, target
Assembled from two labeling passes (GPT-4.1-mini: 7,232 rows; Grok-3-mini: 2… See the full description on the dataset page: https://huggingface.co/datasets/Mahesh111000/hanabi-fireworks-state-tracking.OVIBenchfireworks-dataset
🎆 Fireworks Audio Dataset
A curated dataset of fireworks sounds, built by merging personal field recordings with filtered clips from FSD50K.Designed for audio classification, sound event detection, and edge AI deployment.
📦 Dataset at a glance
Property
Value
Label
fireworks
Format
WAV PCM 16-bit
Sample rate
44 100 Hz
Channels
Stereo (2)
Clip duration
5 s (fixed)
Splits
train (90 %) · test (10 %)
🗂️ Sources
Source… See the full description on the dataset page: https://huggingface.co/datasets/Newton2676/fireworks-dataset.four-digits-multiply-four-digitfour-digits-multiply-open-instructnexus_parallel_functions
Nexus Function Definitions
This dataset reformats function specs from the Nexusflow/Function_Call_Definitions VT_Multi subset.
The function defitions are formatted as JSON Schema objects making them suitable for use with OpenAI compatible APIs.
Dataset Fields
function:str - function spec in JSON Schema syntax serialized as a string with 4-space indent.
Usage
from datasets import load_dataset
dataset = load_dataset("fireworks-ai/nexus_parallel_functions")
llava-pretrain-laionthree-digits-multiply-open-instructfour-digits-multiply-cot-sftnexus_parallel_messages
Nexus Parallel Messages
A parallel function calling validation dataset derived from Nexusflow/VirusTotalMultiple.
The original dataset includes a mix of parallel and nested calls. Here we focus on the former (parallel calls). If the instruction requires sequential calling,
we will generate the initial set of calls only.
For instance, consider the following question:
What is the majority vote from the votes returned by IP address '192.168.1.1'?
The ideal set of calls is… See the full description on the dataset page: https://huggingface.co/datasets/fireworks-ai/nexus_parallel_messages.four-digits-multiply-three-digitllava-instruct-finetuning-mix-688k-ossdrive-thru-synthetic-datasetfireworks-inversionlong-lmsys-chatllm_k12SCAN