datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
nanobubbleeval
NanoBubbleEval v1.0
⚠ For NeurIPS reviewers — use this Croissant URL
Please do NOT use the URL exposed by the "Use this dataset → Croissant"
button at the top-right of this page. That URL triggers a known bug in
mlcroissant==1.0.16 (the version pinned by the
NeurIPS Croissant validator Space)
and produces a FilterFiles error that does not reflect a problem with the
dataset itself.
Use this URL instead — copy the line below verbatim into the validator's
"URL Input" tab:… See the full description on the dataset page: https://huggingface.co/datasets/EliasHossain/nanobubbleeval.nanochat-depo-capability-data
Nanochat Depo Capability Pilot
This dataset is a deterministic natural-language rendering of the Depo directed-cycle
successor task. Each row contains shuffled operational records, one exact multi-hop
question, and its answer. Latent worlds are generated programmatically; no rows were
written or labeled by a language model.
Splits
Split
Worlds
Queries per world
Rows
Renderer family
train
32,768
4
131,072
incident handoff, six structural styles… See the full description on the dataset page: https://huggingface.co/datasets/SolidSnake123/nanochat-depo-capability-data.nanochat-depo-retrieval-copy1-20260715
Nanochat Depo retrieval v1
Each latent 16-node graph yields eight independent, token-aligned, depth-one
query documents. This arm exposes 1 nested edge(s) per
document. Only the answer is supervised in every document; the terminal token
is supervised only for query ordinal 7. This source is separate from and does
not alter Depo-L0 v1.
BCE-Prettybird-Nano-Kayra-v0.1
BCE-Prettybird-Nano-Kayra-v0.1 - 200 AI Brain Mechanism Chat
Kayra is an experimental 200-sample chat dataset developed by PROMETECH A.Ş. for research on Behavioral Consciousness Engine-style control systems. The dataset was synthetically generated using Nemotron Super and is designed to go beyond standard conversation data by exposing layered behavioral signals such as trust scoring, risk level, ethical guardrails, ego–superego balance, KPI tracking, cognitive-level analysis… See the full description on the dataset page: https://huggingface.co/datasets/pthinc/BCE-Prettybird-Nano-Kayra-v0.1.nanochat-depo-composition-depth2-w4-retry-20260715
Nanochat Depo composition v1
Each 16-node single-cycle graph yields eight independent one-query documents:
four base starts paired across query depths (1, 2).
This source contains train and validation splits only. Phase depth is
2; the materialized context width is 4.
docvqa-nanochat
DocVQA for Nanochat
Single-page document QA dataset processed for nanochat fine-tuning.
Description
This dataset is derived from pixparse/docvqa-single-page-questions and has been processed for efficient fine-tuning of small language models with limited context windows.
Modifications from Source
OCR truncation: Answer-priority truncation ensures the answer is always present in the truncated context. Lines containing the answer are prioritized, then surrounding… See the full description on the dataset page: https://huggingface.co/datasets/morgan/docvqa-nanochat.nanochat-depo-l0-symbolic-20260715
Nanochat Depo-L0: symbolic
This is a diagnostic, separately versioned Depo source. Each row contains one
16-node cycle and eight queries at depths 1, 2, 4, and 8. Only the eight
single-letter answers and terminal token are supervised. It is designed for a
one-document-per-sequence training protocol and must not be treated as public
Depo v3 data.
nanochat-world-state-v2-49k-20260714
Nanochat World-State Tracking Capability
Each deterministic latent world describes initial people, rooms, portable objects,
containers, and fixed surfaces followed by a valid chronological event sequence.
The task asks for one exact final location, holder, container, or support. Six
natural renderer styles appear in training; validation and test may use all eight.
The exact state replay engine supplies every answer. No language model generated
or labeled the data. Targeted-v2… See the full description on the dataset page: https://huggingface.co/datasets/SolidSnake123/nanochat-world-state-v2-49k-20260714.nanochat-depo-composition-depth2-w4-transport-20260715
Nanochat Depo composition v1
Each 16-node single-cycle graph yields eight independent one-query documents:
four base starts paired across query depths (1, 2).
This source contains train and validation splits only. Phase depth is
2; the materialized context width is 4.
single-turn-eval-meta_feedback_qwen3-4b_step2_gpt-5-nano_gepa-n32
Single-turn eval — violetxi/meta_feedback_qwen3-4b_step2_gpt-5-nano_gepa
Generated by teaching/inference/single_turn_eval_vllm.py. One row per problem; samples is the list of model responses, scores is per-sample correctness, and mean/best/worst are the aggregates used by mean@N / best@N / worst@N.
Eval results (n_samples_per_example = 32)
Overall
metric
value
n_examples
1006
mean@32
0.1804
best@32
0.3569
worst@32
0.0398
pass_rate… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/single-turn-eval-meta_feedback_qwen3-4b_step2_gpt-5-nano_gepa-n32.nanochat-depo-l0-depth1-curriculum-20260715
Nanochat Depo-L0: symbolic
This is a diagnostic, separately versioned Depo source. Each row contains one
16-node cycle and eight queries under the depth1_only schedule. Only the eight
single-letter answers and terminal token are supervised. It is designed for a
one-document-per-sequence training protocol and must not be treated as public
Depo v3 data.
nanochat-depo-retrieval-width4-20260715
Nanochat Depo retrieval v1
Each latent 16-node graph yields eight independent, token-aligned, depth-one
query documents. This arm exposes 4 nested edge(s) per
document. Only the answer is supervised in every document; the terminal token
is supervised only for query ordinal 7. This source is separate from and does
not alter Depo-L0 v1.
nanochat-depo-composition-depth2-w4-20260715
Nanochat Depo composition v1
Each 16-node single-cycle graph yields eight independent one-query documents:
four base starts paired across query depths (1, 2).
This source contains train and validation splits only. Phase depth is
2; the materialized context width is 4.