hasandmrkrn commited on
Commit
b3a6002
·
verified ·
1 Parent(s): 5208d38

release: publish Kestrel v0.1.0

Browse files

Portable JSON model artifact, minimal model card, and SHA-256 checksum.

Files changed (3) hide show
  1. README.md +70 -0
  2. SHA256SUMS +1 -0
  3. classifier.json +0 -0
README.md ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ pipeline_tag: text-classification
4
+ tags:
5
+ - shell
6
+ - bash
7
+ - cybersecurity
8
+ - safety
9
+ - classifier
10
+ ---
11
+
12
+ # Kestrel
13
+
14
+ Kestrel is a compact local classifier for context-free cyber and system risk in
15
+ individual Bash tool calls.
16
+
17
+ ## Release
18
+
19
+ Version 0.1.0 publishes the portable JSON model artifact used for the Kestrel
20
+ evaluation. This repository does not include training code, training data, or a
21
+ detailed training recipe.
22
+
23
+ | File | Description |
24
+ |---|---|
25
+ | `classifier.json` | Portable Kestrel v0.1.0 model artifact |
26
+ | `SHA256SUMS` | Integrity checksum for the release artifact |
27
+
28
+ The model is custom and is not a Transformers checkpoint. Download the artifact
29
+ with `huggingface_hub`:
30
+
31
+ ```python
32
+ from huggingface_hub import hf_hub_download
33
+
34
+ model_path = hf_hub_download(
35
+ repo_id="kontext-security/Kestrel",
36
+ filename="classifier.json",
37
+ )
38
+ ```
39
+
40
+ ## Evaluation
41
+
42
+ On the published ShellRisk-Bench v0.1 test split, Kestrel reaches 0.947
43
+ precision, 0.922 recall, and 0.934 F1. The test split is a same-source,
44
+ in-distribution holdout; it is not evidence of transfer to a novel command
45
+ dialect.
46
+
47
+ See [ShellRisk-Bench](https://github.com/kontext-security/shellrisk-bench) for
48
+ the fixed split, evaluation methodology, prompts, scorer, aggregate results,
49
+ and per-example Kestrel verdicts. The dataset is available at
50
+ [kontext-security/ShellRisk-Bench](https://huggingface.co/datasets/kontext-security/ShellRisk-Bench).
51
+
52
+ ## Scope and limitations
53
+
54
+ Kestrel evaluates one submitted Bash command without user, task, or session
55
+ context. It does not infer intent or make a complete authorization decision.
56
+ It is intended as one signal layered underneath deterministic protections and
57
+ runtime authorization policy, not as the sole control for executing commands.
58
+
59
+ ## Integrity
60
+
61
+ ```text
62
+ SHA-256 1df8b3e5f2bfc4e1fe95230ee9b3d37f63aaa461a8551982fcc0dd7103c8221b classifier.json
63
+ ```
64
+
65
+ ## License
66
+
67
+ The artifact is published with `license: other`. The public benchmark sources
68
+ retain their own upstream terms; review the
69
+ [dataset provenance](https://github.com/kontext-security/shellrisk-bench/blob/main/DATASETS.md)
70
+ before reuse or redistribution.
SHA256SUMS ADDED
@@ -0,0 +1 @@
 
 
1
+ 1df8b3e5f2bfc4e1fe95230ee9b3d37f63aaa461a8551982fcc0dd7103c8221b classifier.json
classifier.json ADDED
The diff for this file is too large to render. See raw diff