datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
OpenCodeInstruct
OpenCodeInstruct: A Large-scale Instruction Tuning Dataset for Code LLMs
Dataset Description
We introduce OpenCodeInstruct, the largest open-access instruction tuning dataset, comprising 5 million diverse samples. OpenCodeInstruct is designed for supervised fine-tuning (SFT).
Technical Report - Discover the methodology and technical details behind OpenCodeInstruct.
Github Repo - Access the complete pipeline used to perform SFT.
This dataset is ready for… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/OpenCodeInstruct.OpenCodeReasoning
OpenCodeReasoning: Advancing Data Distillation for Competitive Coding
Data Overview
OpenCodeReasoning is the largest reasoning-based synthetic dataset to date for coding, comprises 735,255 samples in Python across 28,319 unique competitive programming
questions. OpenCodeReasoning is designed for supervised fine-tuning (SFT).
Technical Report - Discover the methodology and technical details behind OpenCodeReasoning.
Github Repo - Access the complete pipeline used to… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/OpenCodeReasoning.opencodeinstruct-curatednvidia-OpenCodeInstruct-refined
nvidia-OpenCodeInstruct-refined
A strictly quality-filtered subset of nvidia/OpenCodeInstruct (5M examples). This is a strict subset of EER6/nvidia-OpenCodeInstruct-broad.
Filtering criteria
Both conditions must be satisfied:
Criterion
Threshold
LLM judge min score
= 5 (out of 5)
Unit test pass rate (average_test_score)
= 1.0
LLM judge min score is the minimum across all three dimensions in the llm_judgement field:
requirement_conformance — does the… See the full description on the dataset page: https://huggingface.co/datasets/EER6/nvidia-OpenCodeInstruct-refined.opc-sft-stage2
OpenCoder Dataset
The OpenCoder dataset is composed of the following datasets:
opc-sft-stage1: the sft data used for opencoder sft-stage1
opc-sft-stage2: the sft data used for opencoder sft-stage2 <-- you are here
opc-annealing-corpus: the synthetic data & algorithmic corpus used for opencoder annealing
opc-fineweb-code-corpus: the code-related page recalled from fineweb
opc-fineweb-math-corpus: the math-related page recalled from finewebrefineCode-code-corpus-meta: the meta-data… See the full description on the dataset page: https://huggingface.co/datasets/OpenCoder-LLM/opc-sft-stage2.OpenCodeReasoning-2
OpenCodeReasoning-2: A Large-scale Dataset for Reasoning in Code Generation and Critique
Dataset Description
OpenCodeReasoning-2 is the largest reasoning-based synthetic dataset to date for coding, comprising 1.4M samples in Python and 1.1M samples in C++ across 34,799 unique competitive programming questions.
OpenCodeReasoning-2 is designed for supervised fine-tuning (SFT) tasks of code completion and code critique.
Github Repo - Access the complete pipeline used to… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/OpenCodeReasoning-2.opc-fineweb-code-corpus
OpenCoder Dataset
The OpenCoder dataset is composed of the following datasets:
opc-sft-stage1: the sft data used for opencoder sft-stage1
opc-sft-stage2: the sft data used for opencoder sft-stage2
opc-annealing-corpus: the synthetic data & algorithmic corpus used for opencoder annealing
opc-fineweb-code-corpus: the code-related page recalled from fineweb <-- you are here
opc-fineweb-math-corpus: the math-related page recalled from finewebrefineCode-code-corpus-meta: the meta-data… See the full description on the dataset page: https://huggingface.co/datasets/OpenCoder-LLM/opc-fineweb-code-corpus.opc-sft-stage1
OpenCoder Dataset
The OpenCoder dataset is composed of the following datasets:
opc-sft-stage1: the sft data used for opencoder sft-stage1 <-- you are here
opc-sft-stage2: the sft data used for opencoder sft-stage2
opc-annealing-corpus: the synthetic data & algorithmic corpus used for opencoder annealing
opc-fineweb-code-corpus: the code-related page recalled from fineweb
opc-fineweb-math-corpus: the math-related page recalled from finewebrefineCode-code-corpus-meta: the meta-data… See the full description on the dataset page: https://huggingface.co/datasets/OpenCoder-LLM/opc-sft-stage1.RefineCode-code-corpus-metaThis dataset consists of meta information (including the repository name and file path) of the raw code data from RefineCode. You can collect those files referring to this metadata and reproduce RefineCode!
Note: Currently, we have uploaded the meta data covered by The Stack V2 (About 50% file volume). Due to complex legal considerations, we are unable to provide the complete source code currently. We are working hard to make the remaining part available.
RefineCode is a high-quality… See the full description on the dataset page: https://huggingface.co/datasets/OpenCoder-LLM/RefineCode-code-corpus-meta.OpenCodeReasoning_messages
This is a transformation of the nvidia/OpenCodeReasoning dataset into a format that is more easily digestible by trainers.
OpenCodeReasoning: Advancing Data Distillation for Competitive Coding
Data Overview
OpenCodeReasoning is the largest reasoning-based synthetic dataset to date for coding, comprises 735,255 samples in Python across 28,319 unique competitive programming
questions. OpenCodeReasoning is designed for supervised fine-tuning (SFT).
Technical Report -… See the full description on the dataset page: https://huggingface.co/datasets/sealad886/OpenCodeReasoning_messages.OpenCodeReasoning-2
OpenCodeReasoning-2: A Large-scale Dataset for Reasoning in Code Generation and Critique
Dataset Description
OpenCodeReasoning-2 is the largest reasoning-based synthetic dataset to date for coding, comprising 1.4M samples in Python and 1.1M samples in C++ across 34,799 unique competitive programming questions.
OpenCodeReasoning-2 is designed for supervised fine-tuning (SFT) tasks of code completion and code critique.
Github Repo - Access the complete pipeline used to… See the full description on the dataset page: https://huggingface.co/datasets/cublya/OpenCodeReasoning-2.OpenCodeInstruct-Clean
OpenCodeInstruct Clean
High-quality Python code generation dataset with duplication markers and complexity metrics.
Derived from nvidia/OpenCodeInstruct
after applying strict quality gates.
Quick Stats
Metric
Value
Total rows
388,629
Columns
54
Python-parsable
100.0%
Overview
This dataset contains 388,629 high-quality Python code generation examples
extracted from the nvidia/OpenCodeInstruct corpus.
Each row has been… See the full description on the dataset page: https://huggingface.co/datasets/JetBrains-Research/OpenCodeInstruct-Clean.OpenCodeGeneticInstruct
OpenCodeGeneticInstruct: A large-scale dataset of coding instructions for improving the code generation capabilities of LLMs
Data Overview
OpenCodeGeneticInstruct comprises more than 15M coding instructions in python which is generated synthetically with the Genetic-Instruct [1] approach.
This dataset can be used for supervised fine-tuning (SFT) of LLMs to improve their code genearation capability.
Each sample includes a coding question/instruction and its corrsponding… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/OpenCodeGeneticInstruct.OpenCodeInstruct
OpenCodeInstruct: A Large-scale Instruction Tuning Dataset for Code LLMs
Dataset Description
We introduce OpenCodeInstruct, the largest open-access instruction tuning dataset, comprising 5 million diverse samples. OpenCodeInstruct is designed for supervised fine-tuning (SFT).
Technical Report - Discover the methodology and technical details behind OpenCodeInstruct.
Github Repo - Access the complete pipeline used to perform SFT.
This dataset is ready for… See the full description on the dataset page: https://huggingface.co/datasets/Jeremydh911/OpenCodeInstruct.opc-fineweb-math-corpus
OpenCoder Dataset
The OpenCoder dataset is composed of the following datasets:
opc-sft-stage1: the sft data used for opencoder sft-stage1
opc-sft-stage2: the sft data used for opencoder sft-stage2
opc-annealing-corpus: the synthetic data & algorithmic corpus used for opencoder annealing
opc-fineweb-code-corpus: the code-related page recalled from fineweb
opc-fineweb-math-corpus: the math-related page recalled from fineweb <-- you are here
refineCode-code-corpus-meta: the… See the full description on the dataset page: https://huggingface.co/datasets/OpenCoder-LLM/opc-fineweb-math-corpus.NMT-opencode
OpenCodeReasoning: Advancing Data Distillation for Competitive Coding
Data Overview
OpenCodeReasoning is the largest reasoning-based synthetic dataset to date for coding, comprises 735,255 samples in Python across 28,319 unique competitive programming
questions. OpenCodeReasoning is designed for supervised fine-tuning (SFT).
Technical Report - Discover the methodology and technical details behind OpenCodeReasoning.
Github Repo - Access the complete pipeline used to… See the full description on the dataset page: https://huggingface.co/datasets/Compumacy/NMT-opencode.llm-verifier-freelancer-qwen3.5-122b-131k-opencode-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/llm-verifier-freelancer-qwen3.5-122b-131k-opencode-traces.nemotron-gym-instruction-following-structured-qwen3.5-122b-131k-opencode-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/nemotron-gym-instruction-following-structured-qwen3.5-122b-131k-opencode-traces.nvidia-OpenCodeReasoningopencode_reasoning_filtered
🧠 OpenCode Reasoning (Filtered)
Author: Parvesh Rawal — XenArcAILicense: Inherits from NVIDIA OpenCodeReasoningVersion: Filtered & Structured VariantTotal Examples: 567,850Total Size: 9GB (compressed)
🔍 Overview
This dataset is a curated and cleaned version of split_0 from nvidia/OpenCodeReasoning, optimized for code-level reasoning tasks and instruction tuning.
It’s designed to enhance logic understanding and multistep problem solving for LLMs.
📁 Features… See the full description on the dataset page: https://huggingface.co/datasets/Parveshiiii/opencode_reasoning_filtered.stackexchange-tezos-sandboxes-verified-qwen3.5-122b-131k-opencode-literal-rescue-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/stackexchange-tezos-sandboxes-verified-qwen3.5-122b-131k-opencode-literal-rescue-traces.stackexchange-superuser-sandboxes-verified-qwen3.5-122b-131k-opencode-literal-rescue-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/stackexchange-superuser-sandboxes-verified-qwen3.5-122b-131k-opencode-literal-rescue-traces.OpenCodeReasoning_ShareGPT
Added cnversations column in ShareGPT format
Original README from nvidia/OpenCodeReasoning
OpenCodeReasoning: Advancing Data Distillation for Competitive Coding
Data Overview
OpenCodeReasoning is the largest reasoning-based synthetic dataset to date for coding, comprises 735,255 samples in Python across 28,319 unique competitive programming
questions. OpenCodeReasoning is designed for supervised fine-tuning (SFT).
Technical Report - Discover the methodology and… See the full description on the dataset page: https://huggingface.co/datasets/MaziyarPanahi/OpenCodeReasoning_ShareGPT.stackexchange-overflow-sandboxes-verified-qwen3.5-122b-131k-opencode-literal-rescue-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/stackexchange-overflow-sandboxes-verified-qwen3.5-122b-131k-opencode-literal-rescue-traces.nemotron-gym-identity-following-v2-qwen3.5-122b-131k-opencode-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/nemotron-gym-identity-following-v2-qwen3.5-122b-131k-opencode-traces.exp_rpt_pymethods2test-large-qwen3.5-122b-131k-opencode-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/exp_rpt_pymethods2test-large-qwen3.5-122b-131k-opencode-traces.OpenCodeInstruct-MBPP-Textnemotron-gym-competitive-coding-qwen3.5-122b-131k-opencode-traces
Agent trace dataset
Decoding the literal token IDs
The prompt_token_ids / completion_token_ids / logprobs columns are the
verbatim tokens the serving engine emitted, stored PER AGENT STEP as a
list-of-lists (one inner list per turn). To turn them back into text you MUST
use the exact tokenizer the model was served with — a generic same-family
tokenizer will decode word tokens to garbage.
Served model / tokenizer source: Qwen/Qwen3.5-122B-A10B-FP8
from transformers… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/nemotron-gym-competitive-coding-qwen3.5-122b-131k-opencode-traces.Qwen3-Coder-Next-OpenCode-Preference
Dataset Card — OpenCode Rejection Sampling (Preference)
Overview
This dataset contains 10,920 preference pairs for preference-based training (DPO, KTO, SimPO, ORPO, etc.) on competitive programming tasks. Each pair consists of:
Chosen: a candidate solution that passes 100% of test cases
Rejected: a candidate solution that fails, with a fine-grained rejection type label
Pairs are produced via rejection sampling with Qwen3-Coder-Next: 8 candidate solutions are… See the full description on the dataset page: https://huggingface.co/datasets/zake7749/Qwen3-Coder-Next-OpenCode-Preference.selfinstruct-naive-sandboxes-2-verified-qwen3.5-122b-131k-opencode-traces