hunter-sft-v2

Supervised fine-tuning data for training vulnerability-hunting code assistants: given a C/C++ function, the model reports a verdict (VULNERABLE or CLEAN), names the CWE identifier, explains precisely what is wrong and where, and describes how to fix it.

This is the second iteration of the dataset. v1 taught models the shape of a hunt report but not the judgment (models predicted VULNERABLE ~50% of the time regardless of input and never emitted CWE identifiers). v2 was rebuilt around four fixes: hard clean negatives (patched versions of vulnerable functions, not random clean code), mandatory CWE slots in every output, explanations grounded in the actual code diff, and a held-out validation split so training can be monitored on discrimination rather than loss alone.

Sources

Source	Rows (train)	Role
CWE-enriched BigVul+PrimeVul merge	6,878	Real-world backbone; clean rows are patched versions of vulnerable functions (9,257 easy-clean rows deliberately dropped)
NIST Juliet test suite	5,090	Synthetic good/bad pairs per CWE for CWE-ID grounding; label-leaking identifiers neutralized, testcases with incidental flaws excluded
BigVul before/after pairs	1,328	Vulnerable function + its upstream fix as a contrastive pair
PrimeVul paired subset	1,428	Vulnerable function + patched counterpart

Dataset Structure

Each line is JSON with a single text field containing a JSON-encoded [system, user, assistant] conversation:

{"text": "[{\"role\": \"system\", \"content\": \"You are a vulnerability hunting agent...\"}, {\"role\": \"user\", \"content\": \"Analyze the following function...\"}, {\"role\": \"assistant\", \"content\": \"**Verdict:** VULNERABLE\\n\\n**CWE:** CWE-264 ...\"}]"}

Every assistant message contains fixed slots:

**Verdict:** VULNERABLE

**CWE:** CWE-264 (Permissions, Privileges, and Access Controls)

**Explanation:** The vulnerable code in `...` is:

DeleteCriticalSection (this->synch_cs); g_free (this->synch_cs);

...explanation quoting the vulnerable lines...

**Fix approach:** Replace it with:

...patched lines...

Clean rows use **Verdict:** CLEAN and **CWE:** None.

Data Splits

Loaded with load_dataset("mbenachour/hunter-sft-v2"):

Split	Rows
train	14,724
validation	776

Dataset Creation

Built with an automated pipeline (build_v2.py): download → normalize to the hunt format → enforce the verdict/CWE slots → drop rows whose diffs aren't localizable → deduplicate by normalized function hash → carve the stratified validation split → validate every row (JSON structure, roles, slot regexes, verdict/CWE agreement).

Eval hygiene: deduplicated internally, zero train/validation overlap, and zero overlap with the colin/PrimeVul test split.

Considerations for Using the Data

• C/C++ only. The hunting behavior may not transfer to other languages without additional data.
• Labels derive from fix commits and synthetic test cases, not manual audit. The hard-negative construction (patched = clean) is strong signal, but residual label noise exists, as with all commit-mined datasets.
• The Juliet portion is synthetic. It is excellent for learning CWE patterns and restraint, but real-world code is messier.
• Intended for SFT research on vulnerability detection. Not a certification of model safety or correctness — evaluate on held-out benchmarks (e.g. SecurityEval, PrimeVul test) before any operational use.

Additional Information

Licensing note: this dataset mixes sources with different terms — PrimeVul (MIT), NIST Juliet (public domain, US government work), BigVul and the merged backbone (license not specified upstream). Verify terms against your use case.

Citation: if you use this dataset, please cite the upstream sources: PrimeVul (Ding et al., ICSE 2025), DiverseVul (Chen et al., RAID 2023), BigVul (Fan et al.), and the NIST SARD Juliet test suite.