Datasets:

rockCO78
/

tract-crosswalk-dataset

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
data_hash: string
fetch_timestamp: string
hubs: struct<002-630: struct<branch_root_id: string, children_ids: list<item: null>, depth: int64, hierarc (... 129128 chars omitted)
  child 0, 002-630: struct<branch_root_id: string, children_ids: list<item: null>, depth: int64, hierarchy_path: string, (... 136 chars omitted)
      child 0, branch_root_id: string
      child 1, children_ids: list<item: null>
          child 0, item: null
      child 2, depth: int64
      child 3, hierarchy_path: string
      child 4, hub_id: string
      child 5, is_leaf: bool
      child 6, name: string
      child 7, parent_id: string
      child 8, related_hub_ids: list<item: null>
          child 0, item: null
      child 9, sibling_hub_ids: list<item: string>
          child 0, item: string
  child 1, 002-801: struct<branch_root_id: string, children_ids: list<item: null>, depth: int64, hierarchy_path: string, (... 136 chars omitted)
      child 0, branch_root_id: string
      child 1, children_ids: list<item: null>
          child 0, item: null
      child 2, depth: int64
      child 3, hierarchy_path: string
      child 4, hub_id: string
      child 5, is_leaf: bool
      child 6, name: string
      child 7, parent_id: string
      child 8, related_hub_ids: list<item: null>
          child 0, item: null
      child 9, sibling_hub_ids: list<item: string>
          child 0, item: string
  child 2, 004-517: struct<branch_root_id: string, children_ids: list<item: string>, depth: int64, hierarchy_p
...
t_id: string
      child 8, related_hub_ids: list<item: null>
          child 0, item: null
      child 9, sibling_hub_ids: list<item: string>
          child 0, item: string
label_space: list<item: string>
  child 0, item: string
roots: list<item: string>
  child 0, item: string
version: string
parameters: struct<method: string, top_k: int64>
  child 0, method: string
  child 1, top_k: int64
created_at: string
candidates: list<item: struct<ai_hub_id: string, cosine_similarity: double, status: string, trad_hub_id: string> (... 1 chars omitted)
  child 0, item: struct<ai_hub_id: string, cosine_similarity: double, status: string, trad_hub_id: string>
      child 0, ai_hub_id: string
      child 1, cosine_similarity: double
      child 2, status: string
      child 3, trad_hub_id: string
similarity_stats: struct<matrix_shape: list<item: int64>, max: double, mean: double, min: double, percentiles: struct< (... 85 chars omitted)
  child 0, matrix_shape: list<item: int64>
      child 0, item: int64
  child 1, max: double
  child 2, mean: double
  child 3, min: double
  child 4, percentiles: struct<25: double, 50: double, 75: double, 90: double, 95: double, 99: double>
      child 0, 25: double
      child 1, 50: double
      child 2, 75: double
      child 3, 90: double
      child 4, 95: double
      child 5, 99: double
  child 5, std: double
counts: struct<accepted: int64, rejected: int64, total: int64>
  child 0, accepted: int64
  child 1, rejected: int64
  child 2, total: int64
to
{'candidates': List({'ai_hub_id': Value('string'), 'cosine_similarity': Value('float64'), 'status': Value('string'), 'trad_hub_id': Value('string')}), 'counts': {'accepted': Value('int64'), 'rejected': Value('int64'), 'total': Value('int64')}, 'created_at': Value('string'), 'parameters': {'method': Value('string'), 'top_k': Value('int64')}, 'similarity_stats': {'matrix_shape': List(Value('int64')), 'max': Value('float64'), 'mean': Value('float64'), 'min': Value('float64'), 'percentiles': {'25': Value('float64'), '50': Value('float64'), '75': Value('float64'), '90': Value('float64'), '95': Value('float64'), '99': Value('float64')}, 'std': Value('float64')}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 478, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2815, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2352, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2377, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              data_hash: string
              fetch_timestamp: string
              hubs: struct<002-630: struct<branch_root_id: string, children_ids: list<item: null>, depth: int64, hierarc (... 129128 chars omitted)
                child 0, 002-630: struct<branch_root_id: string, children_ids: list<item: null>, depth: int64, hierarchy_path: string, (... 136 chars omitted)
                    child 0, branch_root_id: string
                    child 1, children_ids: list<item: null>
                        child 0, item: null
                    child 2, depth: int64
                    child 3, hierarchy_path: string
                    child 4, hub_id: string
                    child 5, is_leaf: bool
                    child 6, name: string
                    child 7, parent_id: string
                    child 8, related_hub_ids: list<item: null>
                        child 0, item: null
                    child 9, sibling_hub_ids: list<item: string>
                        child 0, item: string
                child 1, 002-801: struct<branch_root_id: string, children_ids: list<item: null>, depth: int64, hierarchy_path: string, (... 136 chars omitted)
                    child 0, branch_root_id: string
                    child 1, children_ids: list<item: null>
                        child 0, item: null
                    child 2, depth: int64
                    child 3, hierarchy_path: string
                    child 4, hub_id: string
                    child 5, is_leaf: bool
                    child 6, name: string
                    child 7, parent_id: string
                    child 8, related_hub_ids: list<item: null>
                        child 0, item: null
                    child 9, sibling_hub_ids: list<item: string>
                        child 0, item: string
                child 2, 004-517: struct<branch_root_id: string, children_ids: list<item: string>, depth: int64, hierarchy_p
              ...
              t_id: string
                    child 8, related_hub_ids: list<item: null>
                        child 0, item: null
                    child 9, sibling_hub_ids: list<item: string>
                        child 0, item: string
              label_space: list<item: string>
                child 0, item: string
              roots: list<item: string>
                child 0, item: string
              version: string
              parameters: struct<method: string, top_k: int64>
                child 0, method: string
                child 1, top_k: int64
              created_at: string
              candidates: list<item: struct<ai_hub_id: string, cosine_similarity: double, status: string, trad_hub_id: string> (... 1 chars omitted)
                child 0, item: struct<ai_hub_id: string, cosine_similarity: double, status: string, trad_hub_id: string>
                    child 0, ai_hub_id: string
                    child 1, cosine_similarity: double
                    child 2, status: string
                    child 3, trad_hub_id: string
              similarity_stats: struct<matrix_shape: list<item: int64>, max: double, mean: double, min: double, percentiles: struct< (... 85 chars omitted)
                child 0, matrix_shape: list<item: int64>
                    child 0, item: int64
                child 1, max: double
                child 2, mean: double
                child 3, min: double
                child 4, percentiles: struct<25: double, 50: double, 75: double, 90: double, 95: double, 99: double>
                    child 0, 25: double
                    child 1, 50: double
                    child 2, 75: double
                    child 3, 90: double
                    child 4, 95: double
                    child 5, 99: double
                child 5, std: double
              counts: struct<accepted: int64, rejected: int64, total: int64>
                child 0, accepted: int64
                child 1, rejected: int64
                child 2, total: int64
              to
              {'candidates': List({'ai_hub_id': Value('string'), 'cosine_similarity': Value('float64'), 'status': Value('string'), 'trad_hub_id': Value('string')}), 'counts': {'accepted': Value('int64'), 'rejected': Value('int64'), 'total': Value('int64')}, 'created_at': Value('string'), 'parameters': {'method': Value('string'), 'top_k': Value('int64')}, 'similarity_stats': {'matrix_shape': List(Value('int64')), 'max': Value('float64'), 'mean': Value('float64'), 'min': Value('float64'), 'percentiles': {'25': Value('float64'), '50': Value('float64'), '75': Value('float64'), '90': Value('float64'), '95': Value('float64'), '99': Value('float64')}, 'std': Value('float64')}}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

TRACT Crosswalk Dataset

A human-reviewed crosswalk mapping 5,238 security controls from 31 frameworks to 522 CRE hubs — the first large-scale, multi-framework security taxonomy alignment dataset with expert review and calibrated quality metrics.

What Is This Dataset?

This dataset answers a deceptively hard question: when two security frameworks both talk about "access control" or "data encryption," are they actually talking about the same thing?

Security teams today face a maze of overlapping frameworks — NIST 800-53, ISO 27001, MITRE ATLAS, OWASP Top 10, the EU AI Act, and dozens more. Each defines its own controls using its own terminology. Compliance officers, auditors, and security architects spend thousands of hours manually mapping controls between frameworks to answer questions like:

"We're compliant with ISO 27001 — which NIST 800-53 controls does that cover?"
"The EU AI Act requires X — do we already handle this under our MITRE ATLAS controls?"
"Which frameworks address AI model poisoning, and how do their controls compare?"

TRACT solves this by mapping every control to a shared taxonomy: the Common Requirement Enumeration (CRE). Instead of N-to-N pairwise comparisons between frameworks, each control maps to one or more CRE hubs, making cross-framework comparison trivial.

This dataset contains 5,238 control-to-hub assignments combining:

~4,323 ground truth links from OpenCRE (expert-curated by the CRE project maintainers)
~964 model predictions from a fine-tuned BGE-large-v1.5 bi-encoder, each individually reviewed by a cybersecurity domain expert

Key Concepts (Start Here If You're New)

What is CRE? The Common Requirement Enumeration is a universal taxonomy of security topics. Think of it as a Dewey Decimal System for cybersecurity. It organizes ~522 security concepts ("hubs") into a hierarchy — from broad topics like "Authentication" down to specifics like "Multi-factor Authentication > Time-based OTP."

What is a "hub"? A hub is a single node in the CRE hierarchy representing one security concept. For example, hub 615-663 is "Cryptography" and hub 206-830 is "Input validation." Each hub has a unique ID, a name, and a path showing where it sits in the hierarchy.

What is a "control"? A control is a specific security requirement from a framework. For example, NIST 800-53 control "AC-2" is "Account Management" and ISO 27001 control "A.9.2.1" is "User registration and de-registration." Different frameworks describe similar security concepts using different terminology and granularity.

What is a "crosswalk"? A crosswalk maps controls from one system to equivalent concepts in another. This dataset is a crosswalk that maps controls from 31 different frameworks to CRE hubs. If two controls from different frameworks map to the same CRE hub, they address the same security concept.

What is an "assignment"? Each row in this dataset is an assignment — one control mapped to one CRE hub. A single control may map to multiple hubs (e.g., a control about "encrypted authentication" maps to both "Cryptography" and "Authentication" hubs).

Quick Start

from datasets import load_dataset

ds = load_dataset("rockCO78/tract-crosswalk-dataset")
print(f"{len(ds['train'])} assignments across {ds['train'].unique('framework_id')} frameworks")

# Filter to a specific framework
nist_800_53 = ds["train"].filter(lambda x: x["framework_id"] == "nist_800_53")
print(f"NIST 800-53: {len(nist_800_53)} assignments")

# Get only human-reviewed model predictions
reviewed = ds["train"].filter(lambda x: x["assignment_type"].startswith("model_"))
print(f"Model predictions (reviewed): {len(reviewed)}")

# Find all controls that map to the same hub as a given control
target_hub = ds["train"][0]["hub_id"]
same_hub = ds["train"].filter(lambda x: x["hub_id"] == target_hub)
print(f"Controls sharing hub '{target_hub}': {len(same_hub)}")

Loading as pandas

import pandas as pd
df = pd.read_json("crosswalk_v1.0.jsonl", lines=True)

Loading as raw JSONL

import json
with open("crosswalk_v1.0.jsonl") as f:
    assignments = [json.loads(line) for line in f]

Files in This Dataset

File	Description
`crosswalk_v1.0.jsonl`	Main dataset — one JSON object per line, one assignment per row
`framework_metadata.json`	Per-framework statistics (control counts, assignment counts, coverage type)
`cre_hierarchy_v1.1.json`	Full CRE hub hierarchy tree (522 hubs with parent-child relationships)
`hub_descriptions_v1.0.json`	Hub descriptions and metadata for all 522 CRE hubs
`review_metrics.json`	Detailed review quality metrics (acceptance rates, calibration scores, per-framework breakdown)
`bridge_report.json`	Bridge relationships connecting CRE subtrees (from Phase 2B analysis)
`zenodo_metadata.json`	Metadata for Zenodo DOI registration
`LICENSE`	CC-BY-SA-4.0 license text
`README.md`	This file

Dataset Structure

Each row in crosswalk_v1.0.jsonl represents a single control-to-hub assignment:

Field	Type	Description	Example
`control_id`	string	Unique control identifier (framework-scoped)	`"nist_800_53:AC-2"`
`framework_id`	string	Framework identifier	`"nist_800_53"`
`framework_name`	string	Human-readable framework name	`"NIST SP 800-53"`
`section_id`	string	Control section ID within the framework	`"AC-2"`
`control_title`	string	Control title or name	`"Account Management"`
`hub_id`	string	CRE hub identifier	`"615-663"`
`hub_name`	string	Hub display name	`"Cryptography"`
`hub_path`	string	Full hierarchy path (root > ... > leaf)	`"Root > Technical > Crypto"`
`assignment_type`	string	How this assignment was created (see below)	`"ground_truth_linked"`
`confidence`	float or null	Model's calibrated confidence (0.0-1.0). Null for ground truth and reassigned items	`0.872`
`provenance`	string	Data source identifier	`"opencre_ground_truth"`
`review_status`	string	Review outcome	`"accepted"`
`reviewer_notes`	string or null	Free-text notes from the expert reviewer	`null`

Assignment Types Explained

Every assignment in this dataset has an assignment_type field explaining its origin and quality level:

Type	Count	What It Means	Trust Level
`ground_truth_linked`	varies	Manually curated by OpenCRE project maintainers. These are expert-created links between framework controls and CRE hubs.	Highest
`ground_truth_auto`	varies	Automatically derived via transitive chains (e.g., CAPEC attack pattern -> CWE weakness -> CRE hub). Deterministic, not ML.	High
`model_accepted`	680	TRACT model predicted this mapping and a human expert confirmed it was correct.	High
`model_reassigned`	196	TRACT model predicted a mapping but the expert chose a different (better) hub. The `confidence` field is null because the model's score applied to a different hub.	High (expert-corrected)
`model_rejected`	2	TRACT model predicted a mapping but the expert determined no appropriate hub exists. These are included for completeness but represent failed mappings.	N/A (no valid mapping)

For most use cases, filter out model_rejected:

valid = ds["train"].filter(lambda x: x["assignment_type"] != "model_rejected")

Framework Coverage

This dataset covers 31 security, AI safety, and compliance frameworks:

Framework	Controls	Assignments	Coverage Type
AIUC-1 Standard	132	132	model_prediction
ASVS	277	277	ground_truth
BIML	20	21	ground_truth
CAPEC	349	1799	ground_truth
CSA AI Controls Matrix	243	243	model_prediction
CWE	246	613	ground_truth
Cloud Controls Matrix	29	29	ground_truth
CoSAI Landscape of AI Security Risk Map	55	55	model_prediction
DSOMM	183	214	ground_truth
ENISA	38	66	ground_truth
ETSI	27	35	ground_truth
EU AI Act — Regulation (EU) 2024/1689	126	100	model_prediction
EU GPAI Code of Practice	40	40	model_prediction
ISO 27001	93	93	ground_truth
MITRE ATLAS	202	281	mixed
NIST 800-53	300	300	ground_truth
NIST 800-63	25	79	ground_truth
NIST AI 100-2	28	44	ground_truth
NIST AI 600-1 Generative AI Profile	12	10	model_prediction
NIST AI Risk Management Framework	72	72	model_prediction
NIST SSDF	44	46	ground_truth
OWASP AI Exchange	54	47	ground_truth
OWASP Cheat Sheets	50	391	ground_truth
OWASP GenAI Data Security Risks and Mitigations	21	21	model_prediction
OWASP Proactive Controls	10	76	ground_truth
OWASP Top 10 2021	10	17	ground_truth
OWASP Top 10 for Agentic Applications 2026	10	10	model_prediction
OWASP Top 10 for LLM Applications 2025	10	18	ground_truth
OWASP Top10 for ML	7	10	ground_truth
SAMM	30	30	ground_truth
WSTG	59	118	ground_truth

Coverage types:

ground_truth — All assignments come from expert-curated OpenCRE links. These frameworks have been mapped by the CRE project maintainers.
model_prediction — All assignments are TRACT model predictions, each reviewed by a human expert. These are typically newer AI-specific frameworks not yet in OpenCRE.
mixed — Combines ground truth links with model predictions for controls not covered by OpenCRE.

How It Was Made

This dataset was built through a four-stage pipeline:

Stage 1: Ground Truth Import

We imported 4,323 existing links from the OpenCRE project — a community-maintained mapping of security standards. These cover 26 established frameworks (NIST 800-53, ISO 27001, OWASP, CWE, CAPEC, etc.). Links are either manually curated by domain experts (LinkedTo) or derived through deterministic transitive chains (AutomaticallyLinkedTo, e.g., CAPEC->CWE->CRE).

Stage 2: Model Inference

For 10 frameworks with no OpenCRE coverage (primarily newer AI security frameworks), we ran inference using TRACT — a fine-tuned BGE-large-v1.5 bi-encoder trained on the OpenCRE ground truth data. The model encodes both control text and hub descriptions into a shared 1024-dimensional space, then assigns each control to its nearest hub via calibrated cosine similarity.

Key model details:

Architecture: BGE-large-v1.5 with LoRA adapters, contrastive fine-tuning
Training data: OpenCRE ground truth links (leave-one-framework-out cross-validation)
Performance: hit@1 = 0.531 on held-out frameworks
Confidence: Platt-scaled (temperature calibration) — NOT raw cosine similarity

Stage 3: Expert Human Review

A cybersecurity domain expert reviewed all 878 model predictions (878 real predictions + 20 hidden calibration items). For each prediction, the reviewer:

Read the control's full text to understand its security intent
Evaluated whether the model's suggested CRE hub was semantically appropriate
Made one of three decisions: accept (correct hub), reassign (better hub exists), or reject (no hub fits)

Stage 4: Deduplication and Publication

Where a control appeared in multiple data sources (e.g., both ground truth and model prediction for the same control-hub pair), we kept the highest-authority source: ground truth > model prediction. The final dataset contains one row per unique (control, hub) pair.

Review Methodology and Quality

Overall Review Statistics

Metric	Value
Total predictions reviewed	878
Accepted (model was correct)	680 (77.4%)
Reassigned (expert chose better hub)	196 (22.3%)
Rejected (no appropriate hub)	2 (0.2%)
Review completion	100%

Per-Framework Review Breakdown

Framework	Accepted	Reassigned	Rejected	Acceptance Rate
AIUC-1 Standard	38	92	2	29%
CSA AI Controls Matrix	240	3	0	99%
CoSAI Landscape of AI Security Risk Map	25	30	0	45%
EU AI Act — Regulation (EU) 2024/1689	100	0	0	100%
EU GPAI Code of Practice	28	12	0	70%
MITRE ATLAS	177	18	0	91%
NIST AI 600-1 Generative AI Profile	10	0	0	100%
NIST AI Risk Management Framework	42	30	0	58%
OWASP GenAI Data Security Risks and Mitigations	10	11	0	48%
OWASP Top 10 for Agentic Applications 2026	10	0	0	100%

Calibration Quality Check

To measure reviewer quality, we embedded 20 hidden calibration items among the predictions. These were ground-truth assignments where the correct hub was already known — the reviewer did not know which items were calibration tests.

Agreement rate: 65% (13/20 calibration items agreed with ground truth)
All 7 disagreements were reassignments (the reviewer chose a different hub), not rejections — suggesting legitimate alternative interpretations rather than errors

Confidence Scores and Calibration

The confidence field contains the model's calibrated probability that its hub assignment is correct. These are Platt-scaled (temperature-calibrated) softmax outputs, NOT raw cosine similarities.

Interpreting confidence values:

Range	Meaning	Observed Acceptance Rate
> 0.5 (high confidence)	Model is fairly certain	98.2% (161/164)
<= 0.5 (low confidence)	Model is less certain — review more carefully	72.7% (519/714)
OOD items	Model flags these as outside training distribution	7.7% (2/26)

Important: confidence is null for:

Ground truth assignments (no model was involved)
Reassigned predictions (the model's confidence was for a different hub than the one the expert chose)

Who Should Use This

Compliance teams: "We meet ISO 27001 — what does that buy us for NIST 800-53 compliance?" Filter both frameworks to the same hub_ids and you have your mapping.

Security architects: Building a control library that spans multiple frameworks? Use this dataset as the backbone — it handles the taxonomy alignment so you can focus on implementation.

AI security researchers: Studying how AI-specific frameworks (MITRE ATLAS, NIST AI RMF, EU AI Act) relate to traditional security controls? This is the first dataset to systematically map them to a shared taxonomy.

GRC tool builders: Need a machine-readable crosswalk for your platform? This dataset provides structured JSONL with consistent identifiers across all 31 frameworks.

NLP/ML researchers: Interested in security text classification, taxonomy alignment, or domain-specific fine-tuning? This dataset provides labeled training data with expert-reviewed quality annotations.

Usage Examples

Find controls equivalent to a given control

# "Which controls across all frameworks are equivalent to NIST 800-53 AC-2?"
ac2 = df[df["control_id"] == "nist_800_53:AC-2"]
ac2_hubs = set(ac2["hub_id"])
equivalents = df[df["hub_id"].isin(ac2_hubs) & (df["framework_id"] != "nist_800_53")]
print(equivalents[["framework_name", "section_id", "control_title", "hub_name"]])

Cross-framework coverage gap analysis

# "Which CRE hubs does ISO 27001 cover that MITRE ATLAS doesn't?"
iso_hubs = set(df[df["framework_id"] == "iso_27001"]["hub_id"])
atlas_hubs = set(df[df["framework_id"] == "mitre_atlas"]["hub_id"])
gaps = iso_hubs - atlas_hubs
gap_names = df[df["hub_id"].isin(gaps)][["hub_id", "hub_name"]].drop_duplicates()
print(f"ISO 27001 covers {len(gaps)} hubs that MITRE ATLAS does not:")
print(gap_names.head(20))

Filter by trust level

# Only expert-curated and expert-reviewed assignments
high_trust = df[df["assignment_type"].isin([
    "ground_truth_linked", "model_accepted", "model_reassigned"
])]

AI-specific framework analysis

ai_frameworks = ["mitre_atlas", "nist_ai_rmf", "csa_aicm", "eu_ai_act", "owasp_llm_top10"]
ai_controls = df[df["framework_id"].isin(ai_frameworks)]
print(f"AI framework controls: {len(ai_controls)}")
print(ai_controls.groupby("framework_name")["hub_id"].nunique().sort_values(ascending=False))

Known Limitations

Training distribution bias. The model was trained on existing OpenCRE links, which over-represent traditional IT security (NIST, OWASP, CWE) and under-represent newer AI safety concepts. Predictions for AI-specific frameworks may be less accurate — this is reflected in the higher reassignment rate for frameworks like AIUC-1 and CoSAI.
Text quality variance. Some frameworks provide only short control titles without descriptions (53 controls flagged as text_quality: "low" during export). Model confidence for these controls is less reliable.
Single reviewer. All model predictions were reviewed by one cybersecurity expert. Inter-rater reliability is not measured. The 65% calibration agreement rate suggests some ambiguity in mapping decisions, particularly for AI-specific hubs.
CRE taxonomy completeness. The CRE hub hierarchy may not perfectly cover all security concepts in newer AI frameworks. Some controls may not have an ideal hub match — these appear as model_rejected (only 2 in this dataset).
Temporal snapshot. This dataset reflects the state of frameworks and CRE as of May 2026. Frameworks are updated periodically, and new CRE hubs may be added.
Provenance-priority dedup. When the same (control, hub) pair exists in multiple sources, we keep the highest-priority source. This means a ground truth link always overrides a model prediction, even if the model had additional context.

Related Resources

OpenCRE — The source taxonomy and ground truth links
TRACT Model — The BGE-large-v1.5 bi-encoder used for predictions
CRE Project — The open-source CRE project
TRACT Repository — Source code for the TRACT pipeline

License

This dataset is licensed under CC-BY-SA-4.0. You are free to:

Share — copy and redistribute the material in any medium or format
Adapt — remix, transform, and build upon the material for any purpose, including commercially

Under the following terms:

Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made
ShareAlike — If you remix or build upon the material, you must distribute your contributions under the same license

Citation

@dataset{tract_crosswalk_2026,
  title = {TRACT Crosswalk Dataset: Human-Reviewed Security Framework Alignment via CRE Hub Taxonomy},
  author = {Lambros, Rock},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/rockCO78/tract-crosswalk-dataset},
  note = {5,238 assignments mapping 31 frameworks to 522 CRE hubs, with 878 expert-reviewed model predictions}
}

Downloads last month: 19

Total file size:

11.1 MB