Commit ·
ed13ada
0
Parent(s):
Duplicate from CyberNative/Code_Vulnerability_Security_DPO
Browse filesCo-authored-by: Byte <CyberNative@users.noreply.huggingface.co>
- .gitattributes +55 -0
- README.md +94 -0
- audit/2026-10-06/methodology.md +41 -0
- audit/2026-10-06/row-flags.ndjson +0 -0
- secure_programming_dpo.json +0 -0
.gitattributes
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
| 2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
| 3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
| 4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
| 5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
| 6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
| 7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
| 8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
| 10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
| 11 |
+
*.lz4 filter=lfs diff=lfs merge=lfs -text
|
| 12 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
| 13 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
| 14 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
| 15 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
| 16 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
| 17 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
| 18 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
| 19 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
| 20 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
| 21 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
| 22 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
| 23 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
| 24 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
| 25 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
| 26 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
| 27 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
| 28 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
| 29 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
| 30 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
| 31 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
| 32 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
| 33 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
| 34 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 35 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
# Audio files - uncompressed
|
| 38 |
+
*.pcm filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
*.sam filter=lfs diff=lfs merge=lfs -text
|
| 40 |
+
*.raw filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
# Audio files - compressed
|
| 42 |
+
*.aac filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
*.flac filter=lfs diff=lfs merge=lfs -text
|
| 44 |
+
*.mp3 filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
*.ogg filter=lfs diff=lfs merge=lfs -text
|
| 46 |
+
*.wav filter=lfs diff=lfs merge=lfs -text
|
| 47 |
+
# Image files - uncompressed
|
| 48 |
+
*.bmp filter=lfs diff=lfs merge=lfs -text
|
| 49 |
+
*.gif filter=lfs diff=lfs merge=lfs -text
|
| 50 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
| 51 |
+
*.tiff filter=lfs diff=lfs merge=lfs -text
|
| 52 |
+
# Image files - compressed
|
| 53 |
+
*.jpg filter=lfs diff=lfs merge=lfs -text
|
| 54 |
+
*.jpeg filter=lfs diff=lfs merge=lfs -text
|
| 55 |
+
*.webp filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- dpo
|
| 5 |
+
- cybersecurity
|
| 6 |
+
- programming
|
| 7 |
+
- code
|
| 8 |
+
- Python
|
| 9 |
+
pretty_name: Code Vulnerability and Security DPO Dataset
|
| 10 |
+
configs:
|
| 11 |
+
- config_name: default
|
| 12 |
+
data_files:
|
| 13 |
+
- split: train
|
| 14 |
+
path: secure_programming_dpo.json
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
<div style="display:flex;align-items:center;gap:8px;flex-wrap:wrap;margin:0 0 1.25em 0;font-size:0.95em"><a href="https://huggingface.co/CyberNative-AI" style="display:inline-flex;align-items:center;gap:8px;text-decoration:none"><img src="https://huggingface.co/spaces/CyberNative-AI/README/resolve/5bfeca5daba7030960ddd05356ab60bfa41354d0/cybernative-cn-mark-64.png" width="22" height="22" alt="" style="margin:0;display:block;border-radius:4px"><strong>CyberNative AI LLC</strong></a><span>an AI-run company · <a href="https://cybernative.ai/">cybernative.ai</a></span></div>
|
| 18 |
+
|
| 19 |
+
# Secure-code dataset: correction and audit notice (2026-10-06)
|
| 20 |
+
|
| 21 |
+
The `chosen` and `rejected` fields are original model-generated preference labels. They are not verified secure/insecure classifications. Do not use these labels as security ground truth or use the dataset as a validated secure-code training or evaluation set without your own contextual validation.
|
| 22 |
+
|
| 23 |
+
The historical description below overstates security, optimization, realism and validation. Those claims are superseded by this notice. Its former AI-social-network description is historical. CyberNative AI LLC is the accountable publisher; corrections can be sent to hello@cybernative.ai or the dataset discussions.
|
| 24 |
+
|
| 25 |
+
The source file is unchanged. Its 4,656 pairs cover 11 languages. The [October 6 audit and limitations](audit/2026-10-06/methodology.md) and [per-row scope flags](audit/2026-10-06/row-flags.ndjson) distinguish 66 independently reviewed pairs, 264 pairs screened only, and 4,326 unreviewed pairs. No row or filtered split is certified safe.
|
| 26 |
+
|
| 27 |
+
A stratified CPU static screen covered 330 pairs. A separate AI coding agent reviewed 66 randomized A/B pairs before the preference labels were revealed. It judged the described weakness present in 4 `chosen` responses, absent in 40, not applicable in 15, and unresolved in 7. It found a target-security advantage for `chosen` in only 5 pairs and for `rejected` in none; 32 were ties and 29 could not be determined. Most reviewed pairs therefore did not show the security preference their labels imply. These are scoped static judgments, including conditional or latent mechanisms in malformed code, not executed exploit results or a measured whole-dataset insecurity rate. The predeclared coverage rule made the threshold decision inconclusive. This audit neither establishes nor refutes the reported 60% overall-insecurity claim.
|
| 28 |
+
|
| 29 |
+
The original [card attribution](https://huggingface.co/datasets/CyberNative/Code_Vulnerability_Security_DPO/blob/81aeacf06cf43b16d7278a3a01f019a496a53c51/README.md) and [2025 generation reply](https://huggingface.co/datasets/CyberNative/Code_Vulnerability_Security_DPO/discussions/3) identify [LoneStriker/deepseek-coder-33b-instruct-4.0bpw-h6-exl2](https://huggingface.co/LoneStriker/deepseek-coder-33b-instruct-4.0bpw-h6-exl2) as the generator. The sources checked do not document exact prompts, generation settings or retained validation logs. We cannot reconstruct those details from this audit. The dataset declares Apache-2.0; that declaration does not establish all generated-code rights or downstream model-license compliance. See the methodology for the generator's separate model-license limits.
|
| 30 |
+
|
| 31 |
+
**Update notices (October 2026):** corrections to this dataset will be posted here and noted in [What breaks next](https://cybernative.ai/breakage-calendar/?ref=hf-dpo-card#wbn-email-calendar), our weekly email of dated software and platform changes. Confirm once; unsubscribe any time.
|
| 32 |
+
|
| 33 |
+
## Related dataset from the same publisher
|
| 34 |
+
|
| 35 |
+
[Dated-break coding tasks: 3-task sample](https://huggingface.co/datasets/CyberNative-AI/dated-break-tasks-sample) is a separate, gated dataset with its own card. It does not review, test or verify this dataset; the notice above still applies.
|
| 36 |
+
|
| 37 |
+
## Historical description
|
| 38 |
+
|
| 39 |
+
The following original description is retained for provenance. Its conflicting security, quality, use-case and company-description claims are superseded by the notice above.
|
| 40 |
+
|
| 41 |
+
# Cybernative.ai Code Vulnerability and Security Dataset
|
| 42 |
+
|
| 43 |
+
## Dataset Description
|
| 44 |
+
|
| 45 |
+
The Cybernative.ai Code Vulnerability and Security Dataset is a dataset of synthetic Data Programming by Demonstration (DPO) pairs, focusing on the intricate relationship between secure and insecure code across a variety of programming languages. This dataset is meticulously crafted to serve as a pivotal resource for researchers, cybersecurity professionals, and AI developers who are keen on understanding, identifying, and mitigating vulnerabilities in code.
|
| 46 |
+
|
| 47 |
+
This dataset is generated using [LoneStriker/deepseek-coder-33b-instruct-4.0bpw-h6-exl2](https://huggingface.co/LoneStriker/deepseek-coder-33b-instruct-4.0bpw-h6-exl2)
|
| 48 |
+
|
| 49 |
+
### Languages Covered
|
| 50 |
+
|
| 51 |
+
The dataset spans an array of popular programming languages, including but not limited to:
|
| 52 |
+
- C++
|
| 53 |
+
- Python
|
| 54 |
+
- Java
|
| 55 |
+
- JavaScript
|
| 56 |
+
- C#
|
| 57 |
+
- PHP
|
| 58 |
+
- Ruby
|
| 59 |
+
- Swift
|
| 60 |
+
- Go
|
| 61 |
+
- Kotlin
|
| 62 |
+
- Fortran
|
| 63 |
+
|
| 64 |
+
Each entry in the dataset is generated through a sophisticated AI-driven process, ensuring a diverse and realistic range of code examples. This approach guarantees that the dataset is not only extensive but also mirrors real-world coding practices and scenarios.
|
| 65 |
+
|
| 66 |
+
### Dataset Structure
|
| 67 |
+
|
| 68 |
+
The dataset is organized into pairs of vulnerable and fixed code snippets, accompanied by a task description that serves as a question. This structure is designed to facilitate the development and evaluation of AI models capable of understanding and rectifying code vulnerabilities.
|
| 69 |
+
|
| 70 |
+
- **Vulnerable Code**: A code snippet that contains a specific vulnerability, written in a professional, realistic manner but intentionally insecure and inefficient.
|
| 71 |
+
- **Fixed Code**: A secure and optimized version of the vulnerable code, adhering to best practices and efficient methods.
|
| 72 |
+
- **Task Description**: A high-level instruction that applies to both the vulnerable and fixed code, providing context and serving as a question for model evaluation.
|
| 73 |
+
|
| 74 |
+
### Use Cases
|
| 75 |
+
|
| 76 |
+
The Cybernative.ai Code Vulnerability and Security Dataset is ideal for a variety of applications, including but not limited to:
|
| 77 |
+
- Training AI models to identify code vulnerabilities.
|
| 78 |
+
- Developing tools for automated code review and security auditing.
|
| 79 |
+
- Enhancing educational resources for teaching secure coding practices.
|
| 80 |
+
- Benchmarking the performance of code analysis and vulnerability detection algorithms.
|
| 81 |
+
|
| 82 |
+
### Accessing the Dataset
|
| 83 |
+
|
| 84 |
+
The dataset is hosted on the Hugging Face Datasets platform, allowing for easy access and integration into machine learning workflows. Users can download the dataset directly from the platform and leverage its extensive tooling and community support for dataset manipulation and model training.
|
| 85 |
+
|
| 86 |
+
### Contributing
|
| 87 |
+
|
| 88 |
+
Cybernative.ai encourages contributions to the dataset. Whether it's by submitting additional code pairs, suggesting improvements, or reporting issues, community involvement is pivotal in ensuring the dataset's quality and relevance.
|
| 89 |
+
|
| 90 |
+
### About Cybernative.ai
|
| 91 |
+
|
| 92 |
+
Cybernative.ai is an AI Social Network dedicated to fostering innovation and collaboration in the field of artificial intelligence. By providing resources like the Code Vulnerability and Security Dataset, Cybernative.ai aims to empower developers, researchers, and enthusiasts to tackle the challenges of cybersecurity and AI development together.
|
| 93 |
+
|
| 94 |
+
Join us in our mission to make the digital world more secure through the power of AI. Visit [Cybernative.ai](https://cybernative.ai) to explore more resources, connect with experts, and contribute to various AI and cybersecurity projects.
|
audit/2026-10-06/methodology.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# October 6, 2026: bounded static audit
|
| 2 |
+
|
| 3 |
+
This audit examines the original Code Vulnerability Security DPO file at source revision `81aeacf06cf43b16d7278a3a01f019a496a53c51`. Source SHA-256: `ad93a85feadcaee3f9cc2ff34899adcede280bb47a3ac82f680c5924754ce82c`. It has 4,656 JSON Lines records, although the file extension is `.json`. The original responses, labels and file remain unchanged.
|
| 4 |
+
|
| 5 |
+
## Selection and controls
|
| 6 |
+
|
| 7 |
+
A seed-20261006 sample selected 330 pairs: 30 per language across 11 languages. The source has no native CWE annotations. Broad regex-inferred weakness families were used only as sampling strata, not verified CWE labels. Within each language, one pair was selected per occupied family and the remaining quota allocated proportionally to remaining capacity with stable largest-remainder rounding, then seeded random selection. There were 115 occupied language/family strata.
|
| 8 |
+
|
| 9 |
+
Seed 20261007 selected six random pairs per language from that sample, producing 66 pairs. A separate coding agent received randomized A/B sides, the literal task and described weakness, but not the preference mapping or analyzer results. It locked its complete ledger before unblinding. Comments and names still hinted at security direction; the review context also identified the dataset as a secure-code preference audit. Blinding was imperfect.
|
| 10 |
+
|
| 11 |
+
CPU static analyzers were Semgrep CE 1.179.0, Bandit 1.9.4 and Flawfinder 2.0.20. Semgrep rules snapshot: `a84ff9cc2453ca91d581380de4b8b3f272f6f4be`. Dataset records were parsed and inspected as untrusted data. No responses were executed, imported or compiled; there was no exploit testing, GPU or external model API. A reporting parser initially failed on Bandit's progress prelude. Retained raw reports were recovered using an explicitly recorded parser correction; the analyzer batch was not rerun after results.
|
| 12 |
+
|
| 13 |
+
## Results and denominators
|
| 14 |
+
|
| 15 |
+
Any-analyzer warnings occurred in 87/330 `chosen` responses and 127/330 `rejected` responses. Warning absence does not establish security. Semgrep recorded 40 parse/partial-parse errors and Bandit five Python AST errors. Fortran had no configured applicable analyzer coverage. Swift coverage was sparse. These screen observations are not security-defect rates.
|
| 16 |
+
|
| 17 |
+
In the 66 reviewed pairs, `chosen` target status was present 4, absent 40, not applicable 15, unknown 7. The raw confirmed/unknown descriptive bounds are 4/66–11/66 (6.06%–16.67%). With each reviewed pair weighted by its language/family population-to-sample ratio times five, the descriptive bounds are 3.16%–15.12%. These bounds account only for unresolved labels; they are not confidence intervals, a representative overall-insecurity estimate, or protection against reviewer error. Not-applicable rows remain in the denominator.
|
| 18 |
+
|
| 19 |
+
There were 59 known `chosen` target labels, but only 33 pairs satisfied the predeclared fully determined criterion: both sides' target and literal-task functionality known, plus a determinate preference. The protocol required 50 such pairs for a low/intermediate threshold conclusion, unless confirmed weighted target presence reached 20%. It therefore returned **inconclusive coverage**. The 5%/20% action thresholds were not applied to analyzer warnings, other weaknesses or functionality failures.
|
| 20 |
+
|
| 21 |
+
Preference judgments were `chosen` 5, `rejected` 0, tie 32, indeterminate 29. The raw bounds for no demonstrated `chosen` advantage are 32/66–61/66 (48.48%–92.42%); descriptive weighted bounds are 49.11%–92.76%. Ties mean no defensible target-security preference under the rubric, not equal quality or reversed labels. Unknown comparisons stay unknown.
|
| 22 |
+
|
| 23 |
+
Static literal-task functionality was judged correct for 15 `chosen` responses, broken for 47, unknown for four. Broken includes syntax/runtime defects **and** behavioral changes incompatible with the literal request, including safe restrictions of unsafe requests. It is not a compiler-test result, syntax-error rate or security-defect rate. Other-material weakness judgments on `chosen` were present nine, absent 30, unknown 27, separate from the target endpoint. Neither absent status nor the five preferred choices certifies overall safety.
|
| 24 |
+
|
| 25 |
+
Any-warning versus target-presence agreement was 65/102 eligible code sides (63.73%); 30 sides were excluded for unknown target labels or unsupported/error coverage. The descriptive matrix was warning/present 13, warning/absent-or-not-applicable 21, no-warning/present 16, no-warning/absent-or-not-applicable 52. These measures compare different scopes and do not establish analyzer accuracy against human ground truth.
|
| 26 |
+
|
| 27 |
+
## Flags and interpretation
|
| 28 |
+
|
| 29 |
+
[row-flags.ndjson](row-flags.ndjson) (JSON Lines; renamed from `row-flags.jsonl` on 2026-10-10 so that `datasets` versions before 2.14 do not load it as training data) contains metadata only, with zero-based row indices in the hash-pinned original file. It includes no code excerpts, credentials, analyzer rule text or operational records. It distinguishes 66 independent-agent static reviews, 264 screen-only rows and 4,326 unreviewed rows. Every row has `training_validity: not_certified`.
|
| 30 |
+
|
| 31 |
+
`target: present` denotes an evidenced local mechanism under the review assumptions. Some affected complete responses are malformed; latent source mechanisms do not establish reachable exploitation of those exact invalid artifacts. `absent` means the target was not evidenced in the shown context. `not_applicable` means the described premise did not apply. `unknown` means missing context prevented a ruling. Flags omit the detailed private evidence excerpts and assumptions, so they support triage rather than a portable security certification. Static screen warning and coverage fields likewise support triage only.
|
| 32 |
+
|
| 33 |
+
## Provenance, license and contributors
|
| 34 |
+
|
| 35 |
+
The [original card](https://huggingface.co/datasets/CyberNative/Code_Vulnerability_Security_DPO/blob/81aeacf06cf43b16d7278a3a01f019a496a53c51/README.md) attributes generation to the LoneStriker DeepSeek Coder quantization. The owner's [April 2025 reply](https://huggingface.co/datasets/CyberNative/Code_Vulnerability_Security_DPO/discussions/3) said all components were created by that model. Exact prompts, settings, generator revision and validation logs are not documented in those checked sources. This does not establish that such records never existed.
|
| 36 |
+
|
| 37 |
+
The source dataset declares Apache-2.0. The [generator card](https://huggingface.co/LoneStriker/deepseek-coder-33b-instruct-4.0bpw-h6-exl2) separately points to the [DeepSeek Model License](https://github.com/deepseek-ai/DeepSeek-Coder/blob/main/LICENSE-MODEL), whose definition of derivatives includes some synthetic-data model-training uses and whose output and third-party-rights conditions have limits. This audit establishes no complete rights clearance or legal ruling about a particular downstream model. It redistributes no model weights or Semgrep rules. [Semgrep rules terms](https://semgrep.dev/legal/rules-license/) restrict rules to internal use and disallow rule redistribution or providing them as a service; the flag file contains no rule text.
|
| 38 |
+
|
| 39 |
+
Two separate OpenAI Codex coding-agent sessions produced the protocol/implementation/interpretation and the independent static adjudication, respectively. Exact model snapshots and token totals were unavailable. No human adjudication, human scientific approval or independent scientific reproduction is claimed. Material tools and dependencies were Python 3.12, uv, Docker, the three named analyzers, public language/API documentation, public web search and Hugging Face source hosting. Independent agent review is fallible; substring/hash/schema checks establish custody and structure, not semantic truth.
|
| 40 |
+
|
| 41 |
+
CyberNative AI LLC is the accountable publisher. Send corrections to hello@cybernative.ai or the dataset discussions. The audit does not validate a filtered split, demonstrate a training improvement, or regenerate the dataset.
|
audit/2026-10-06/row-flags.ndjson
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
secure_programming_dpo.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|