datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Ornith-1.0-9B-atlas
juiceb0xc0de/Ornith-1.0-9B-atlas
A brain atlas for deepreinforce-ai/Ornith-1.0-9B, the 9B agentic-coding model that reports SOTA results on Terminal-Bench, SWE-Bench, and other agentic coding benchmarks. This is not a chat dataset or a benchmark — it is an internal-mechanics map built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.
If you want to know why this model survives surgical edits, where… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/Ornith-1.0-9B-atlas.Ornith-1.5-35B-A3B-GGUF-metricsCUDA-L2
CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning
🥳 Introduction
CUDA-L2 is a system that combines large language models (LLMs) and reinforcement learning (RL) to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. CUDA-L2 systematically outperforms major matmul baselines to date, from the widely-used torch.matmul to state-of-the-art NVIDIA closed-source libraries (cuBLAS… See the full description on the dataset page: https://huggingface.co/datasets/ornith-ai/CUDA-L2.Ornith-1.5-9B-GGUF-metricsornithoscopeOrnithoscope dataset is the dataset used to train the model for the Ornithoscope project.molt-ornith-eval
molt — a behavioral eval harness for extreme MoE quantization (Ornith-1.0-397B @ 2.41 bpw)
molt is an autoresearch-style ratchet loop that
searches quantization recipes and serving configs for deepreinforce-ai/Ornith-1.0-397B
(512-expert qwen3.5 MoE) until it serves well on one specific consumer box: 2×RTX 4090,
90 GB DDR5, Gen5 SSD. This dataset is the frozen evaluation harness, golden references, the
research journal, and the documentation — published for community feedback.… See the full description on the dataset page: https://huggingface.co/datasets/SEBK4C/molt-ornith-eval.CUDA-L1
| 🏠 Project Page | 📄 Paper | 🔥 Demo
🥳 Introduction
In this paper, we introduce CUDA-L1, an automated reinforcement learning (RL) framework for CUDA optimization. The core of CUDA-L1 is a contrastive RL model, a newly-designed RL system to enhance optimization through comparative learning.… See the full description on the dataset page: https://huggingface.co/datasets/ornith-ai/CUDA-L1.Ornith-1.5-35B-A3B-Nemotron-v2-100M
Ornith 1.5 35B A3B Nemotron v2 100M
This dataset contains 108,729 English conversations with
108,729 regenerated assistant turns and 100,014,884 generated
assistant completion tokens. 100M refers to the completion-token target, not the
number of examples.
The prompt mix is a deterministic sample from
nvidia/Nemotron-Post-Training-Dataset-v2.
It covers the source dataset's chat, code, math, and STEM subsets. Every assistant turn
was regenerated with ornith-ai/Ornith-1.5-35B-A3B;… See the full description on the dataset page: https://huggingface.co/datasets/jzinno/Ornith-1.5-35B-A3B-Nemotron-v2-100M.ornith-35b-agentic-eval
Ornith-35B agentic evaluation harness + full journals
The complete, reproducible evidence behind
SEBK4C/Ornith-1.0-35B-MTP-GGUF:
a single day of controlled head-to-head experiments on locally-served
Ornith-1.0-35B quantizations (plus the 397B 2.41bpw research quant), all
machine-scored — no LLM judge anywhere.
What's here
harness/ tasks.py 24-task base battery (agentic tools / code+hidden-tests /
exact-answer reasoning /… See the full description on the dataset page: https://huggingface.co/datasets/SEBK4C/ornith-35b-agentic-eval.ornith-256k-bench
ornith-256k-bench: Raw measurement records for LocalMoE (Ornith-1.5-35B-A3B on M4 Max)
Raw JSON-lines/JSON records backing the LocalMoE study: serving Ornith-1.5
35B-A3B (3B active). The original Q4_K_M experiment and the later Q6_K,
reasoning-on comparison are both retained. The Q4_K_M long-context proof records
a 260,013-prompt-token request on an Apple M4 Max (64 GB); the Q6_K record is a
separate controlled short-workload comparison against a dense 27B + DFlash2 baseline.… See the full description on the dataset page: https://huggingface.co/datasets/ChengyiX/ornith-256k-bench.ornith-coder-sft-v1
ORNITH-CODER-SFT-V1
Coding-specialized SFT dataset.
Target domains
math: 25,000
tool_use: 17,000
general_coding: 20,000
agentic_coding: 30,000
reasoning: 15,000
Maximum target: 107,000 examples.
Ingestion
The pipeline does not use datasets.load_dataset().
Parquet sources are processed with PyArrow batches.
Large JSONL sources are streamed directly over HTTP.
Global deduplication uses SQLite and SHA-256 content hashes.
Hard limits… See the full description on the dataset page: https://huggingface.co/datasets/usernamebetter/ornith-coder-sft-v1.ornith-coder-sft-v2
ORNITH-CODER-SFT-V1
Coding-specialized SFT dataset.
Domain targets
math: 25,000
tool_use: 17,000
general_coding: 20,000
agentic_coding: 30,000
reasoning: 15,000
Maximum target: 107,000 examples.
Ingestion
No datasets.load_dataset()
Direct PyArrow Parquet processing
Direct remote HTTP JSONL streaming
SQLite SHA-256 global deduplication
Source-aware schema normalization
Hard limits
Maximum Parquet file: 1.07 GB
Maximum Parquet… See the full description on the dataset page: https://huggingface.co/datasets/usernamebetter/ornith-coder-sft-v2.Dr.Sparse-Ornith15-9B-eval-b200-otf-spgemm-partial
Dr.Sparse — Ornith-1.5-9B SpGEMM baseline (partial, 12/81 matrices)
Partial baseline of ornith-ai/Ornith-1.5-9B on the Dr.Sparse OTF held-out test set,
SpGEMM only, levels 1-3 (level4 excluded). B200, single trajectory (no tree search).
Why this run is partial
The run was stopped after 12 of 81 matrices. HiPerGator terminates jobs that hold a GPU
without using it, and this eval layout gives each matrix its own GPU while the agent spends
most of each iteration… See the full description on the dataset page: https://huggingface.co/datasets/DiogenesChen122/Dr.Sparse-Ornith15-9B-eval-b200-otf-spgemm-partial.ornith-android-instruct
Ornith Android Instruct
5,500 instruction-tuning examples for Android development — Kotlin-first, Jetpack Compose-first, current (non-deprecated) APIs. Built as the training corpus for Ornith, a small language model specialized in writing Android application code, and usable for fine-tuning any code LLM toward modern Android development.
Train: 5,225 examples (ornith_dataset_merged.jsonl)
Eval: 275 held-out examples (ornith_eval_merged.jsonl)
Format: JSONL, one example per line… See the full description on the dataset page: https://huggingface.co/datasets/giggiovpg/ornith-android-instruct.humatheque-vlm-pred-ornith15-9bimnet1k_platypus_duckbill_duckbilled_platypus_duck-billed_platypus_Ornithorhynchus_anatinusimnet1k_platypus_duckbill_duckbilled_platypus_duck-billed_platypus_Ornithorhynchus_anati