File size: 4,975 Bytes
abefb0c
e36ee3e
abefb0c
 
 
 
caa9604
abefb0c
caa9604
 
 
abefb0c
caa9604
abefb0c
caa9604
abefb0c
caa9604
abefb0c
 
 
c16d2d9
 
 
 
abefb0c
 
 
 
 
42485f3
 
 
 
 
 
 
 
 
abefb0c
42485f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
---
task_type: figma_to_code
dataset_info:
  features:
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: tags
    list: string
  - name: description
    dtype: string
  - name: figma_data
    dtype: string
  - name: access_level
    dtype: string
  - name: created_at
    dtype: string
  splits:
  - name: train
    num_bytes: 64990
    num_examples: 37
  download_size: 35843
  dataset_size: 64990
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: mit
language:
- en
tags:
- figma-to-code
- autonomous-agents
pretty_name: CREW:Figma-to-Code
size_categories:
- n<1K
---

  
  # CREW: Figma to Code

  A benchmark for evaluating AI coding agents on **Figma-to-code generation** — converting real-world Figma community designs
   into production-ready React + Tailwind CSS applications.

  Each task gives an agent full access to a Figma file via MCP tools. The agent must extract the design system, generate
  components, build successfully, and deploy a live preview. Outputs are evaluated through human preference (ELO) and
  automated verifiers.

  ## Benchmark at a Glance

  | | |
  |---|---|
  | **Tasks** | 37 real Figma community designs |
  | **Agents tested** | Claude Code (Opus 4.6), Codex (GPT-5.2), Gemini CLI (3.1 Pro) |
  | **Total runs** | 96 autonomous agent executions |
  | **Human evaluations** | 135 pairwise preference votes |
  | **Primary metric** | Human Preference ELO (Bradley-Terry) |
  | **Task duration** | 6–40 expert-hours equivalent per task |

  ## Leaderboard (Human Preference ELO)

  | Rank | Agent | Model | ELO | 95% CI | Win % |
  |------|-------|-------|-----|--------|-------|
  | 1 | Codex | GPT-5.2 | 1054 | [1005, 1114] | 56.8% |
  | 2 | Claude Code | Opus 4.6 | 1039 | [987, 1093] | 55.1% |
  | 3 | Gemini CLI | 3.1 Pro | 907 | [842, 965] | 31.1% |

  Top two agents are statistically indistinguishable (p=0.67, Cohen's h=0.08); both significantly outperform Gemini CLI
  (p<0.05).

  Live leaderboard: [evals.metaphi.ai](https://evals.metaphi.ai)

  ## Dataset Schema

  Each row represents one Figma design task:

  | Column | Type | Description |
  |--------|------|-------------|
  | `id` | string | Unique task identifier |
  | `figma_data.figma_file_key` | string | Figma file ID for API access |
  | `figma_data.figma_file_url` | string | Full Figma URL |
  | `figma_data.figma_file_name` | string | Design name/title |
  | `figma_data.is_site` | bool | Whether design is a Figma Site |
  | `description` | string | Design context and task description |

  ## Task Scenarios

  Tasks span 7 complexity levels, from single-component extraction to full multi-page applications:

  1. **E-commerce Product Page** (12 hrs) — PDP with image gallery, variant selectors, inventory states, cart integration
  2. **Mobile Onboarding Flow** (16 hrs) — Multi-step flow with transitions, conditional branching, state management
  3. **Component Set with States** (6 hrs) — Variant matrix extraction, typed props, conditional rendering
  4. **Design Tokens to Theme** (8 hrs) — Variables, typography, effects → Tailwind config + CSS custom properties
  5. **Multi-Page Webapp** (40 hrs) — 5+ pages with routing, shared components, consistent theming
  6. **Animation-Heavy Interface** (20 hrs) — Smart Animate → Framer Motion with precise timing choreography
  7. **Messy Enterprise File** (24 hrs) — Real-world chaos: inconsistent naming, duplicates, orphaned components

  ## Data Curation

   Sourced from licensing partnerships with enterprises, domain-experts and community dataset curation.

  ## Evaluation Framework

  | Verifier | Type | Method |
  |----------|------|--------|
  | **Human Preference** | Subjective | Pairwise comparison → Bradley-Terry ELO |
  | **Visual Judge** | Subjective | VLM screenshot comparison (design vs. output) |
  | **Skill Verifier** | Subjective | Task-specific rubrics (build, tokens, components, fidelity) |
  | **Behavior Verifier** | Subjective | Agent trajectory analysis (error recovery, tool usage) |

  ## Agent Error Recovery

  Across 96 runs, agents encountered 590 errors with a 70.3% autonomous recovery rate:

  | Error Type | Count | Recovery |
  |------------|-------|----------|
  | Tool call failure | 419 | 66.3% |
  | Git error | 64 | 75.0% |
  | Syntax error | 33 | 90.9% |
  | Build error | 11 | 100.0% |

  ## Usage

  ```python
  from datasets import load_dataset

  ds = load_dataset("metaphilabs/figma", split="train")

  # Each row contains a Figma file key for API access
  for task in ds:
      print(task["id"], task["figma_data"]["figma_file_name"])

  Citation

  @misc{metaphi2026crew,
    title={CREW: Enterprise Agent Benchmarks},
    author={Metaphi Labs},
    year={2026},
    url={https://evals.metaphi.ai}
  }

  Links

  - Leaderboard: https://evals.metaphi.ai
  - Website: https://metaphi.ai
  - Collection: https://huggingface.co/collections/metaphilabs/crew