contra-labs commited on
Commit
8c98456
·
verified ·
1 Parent(s): 1fa8da0

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +42 -33
README.md CHANGED
@@ -50,7 +50,7 @@ Most AI evaluation treats evaluator disagreement as noise to be averaged away. I
50
 
51
  ## What the dataset contains
52
 
53
- Domain professionals were drawn from a network of independent creatives and matched to the model category most relevant to their workflow, then asked to evaluate anonymized, randomly ordered model outputs at each phase. Prompts chain across phases to mimic a real designer's process: Mockup and Refinement prompts build on the prior phase and supply seed images where applicable.
54
 
55
  | | |
56
  |---|---|
@@ -61,7 +61,8 @@ Domain professionals were drawn from a network of independent creatives and matc
61
  | Models evaluated | 13 |
62
  | Evaluators | 31 (anonymized) |
63
  | Pairwise judgments | 3,174 |
64
- | Scalar evaluation rows | 2,247 (812 with numeric 1-5 ratings — see note) |
 
65
  | Qualitative feedback rows | 2,247 |
66
 
67
  ### Domains, modalities, and models
@@ -80,40 +81,46 @@ Domain professionals were drawn from a network of independent creatives and matc
80
  - **Mockup** — the chosen direction is realized: product shots, scene composition, brand identity.
81
  - **Refinement** — near production-ready; targeted edits for consistency and polish.
82
 
83
- ## Format and schema
 
 
84
 
85
- Five CSV files, joinable on shared keys. All tables join on `prompt_id`; `model_outputs.csv` adds `output_id` (`{prompt_id}__{model_id}`); the scalar and qualitative tables share `evaluation_id` (`{prompt_id}__{model_id}__{user_id}`). `domain` and `stage` can be recovered for any record from `prompts_workflow.csv` or `model_outputs.csv`.
 
 
 
86
 
87
  ```python
88
  from datasets import load_dataset
89
 
90
- pairwise = load_dataset("contra-labs/HCB", "pairwise_comparisons", split="train")
91
- prompts = load_dataset("contra-labs/HCB", "prompts_workflow", split="train")
92
 
93
- row = pairwise[0]
94
- print(row["domain"], row["stage"], "->", row["chosen_model"])
 
95
  ```
96
 
 
 
 
97
  ### `prompts_workflow.csv` — 95 rows (one per prompt)
98
 
99
- | Column | Description |
100
- |---|---|
101
- | `prompt_id` | Unique prompt identifier (e.g. `ad_image_idea_4152`). |
102
  | `domain` | One of the five creative domains. |
103
  | `stage` | `Ideation`, `Mockup`, or `Refinement`. |
104
  | `prompt_text` | The text prompt presented to the models. |
105
- | `input_image` | URL of the seed/reference image used as model input. Present mainly for Mockup and Refinement (44 of 95 rows); empty for most Ideation prompts. |
 
106
 
107
  ### `model_outputs.csv` — 380 rows (one per prompt × model)
108
 
109
- | Column | Description |
110
- |---|---|
111
- | `output_id` | Unique output identifier (`{prompt_id}__{model_id}`). |
112
  | `prompt_id` | Foreign key to `prompts_workflow.csv`. |
113
  | `domain`, `stage` | Creative domain and workflow phase. |
114
  | `model_id` | Model that produced the output. |
115
- | `model_output` | The generated artifact. **Format depends on modality:** image domains (Ad Images, Brand Design) store a `cloudinary:<asset_id>` reference; code domains (Desktop App, Landing Pages) store the raw generated HTML/code inline; the Ad Video field is empty in this release (videos are not bundled in the CSV). |
116
- | `contestant_id` | Internal identifier for the output as presented in the evaluation interface. |
117
 
118
  ### `pairwise_comparisons.csv` — 3,174 rows
119
 
@@ -121,38 +128,40 @@ Forced-choice head-to-head preferences. For each prompt, evaluators compared all
121
 
122
  | Column | Description |
123
  |---|---|
124
- | `comparison_id` | Unique comparison identifier. |
125
  | `prompt_id` | Foreign key to `prompts_workflow.csv`. |
126
  | `domain`, `stage` | Creative domain and workflow phase. |
127
- | `user_id` | Anonymized evaluator identifier. |
128
  | `user_core_skill` | Evaluator's professional specialty (e.g. Brand Designer, Product Designer, Digital Marketer). |
129
  | `model_left`, `model_right` | The two models compared. |
130
  | `chosen_model` | The preferred model (always equal to `model_left` or `model_right`). |
 
131
 
132
- ### `scalar_feedback.csv` — 2,247 rows
 
133
 
134
- 1-5 Likert ratings on three dimensions, ordered from most objective to most taste-driven.
135
 
136
  | Column | Description |
137
  |---|---|
138
- | `evaluation_id` | Unique identifier (`{prompt_id}__{model_id}__{user_id}`). |
139
- | `prompt_id`, `user_id`, `user_core_skill`, `model_id` | Join keys and evaluator/model identifiers. |
140
- | `prompt_adherence` | 1-5: how faithful the output is to the prompt (least subjective). |
141
- | `usability` | 1-5: how well the output would function in a professional context. |
142
- | `visual_appeal` | 1-5: how visually interesting, cohesive, and polished it is (most taste-driven). |
143
 
144
  ### `qualitative_feedback.csv` — 2,247 rows
145
 
146
- Open-ended written rationale, machine-coded into themes. Feedback was stripped of PII and model identities, then passed through a deductive coding pass (GPT-4o against a predefined codebook) to assign themes, per-theme sentiment, and representative quotes.
147
 
148
  | Column | Description |
149
  |---|---|
150
- | `evaluation_id` | Unique identifier (`{prompt_id}__{model_id}__{user_id}`). |
151
- | `prompt_id`, `user_id`, `user_core_skill`, `model_id` | Join keys and evaluator/model identifiers. |
152
- | `model_label` | Blinded label shown to the evaluator (A/B/C/D). |
153
- | `raw_feedback` | The evaluator's free-text response. |
154
- | `assigned_themes` | JSON list of coded themes (e.g. `prompt_adherence`, `visual_hierarchy`, `typography`). |
155
- | `theme_sentiment` | JSON object mapping each theme to `positive` / `negative` / `neutral`. |
 
 
 
156
  | `key_quotes` | JSON object mapping each theme to a supporting quote from the feedback. |
157
 
158
  ## Curation
 
50
 
51
  ## What the dataset contains
52
 
53
+ Domain professionals were drawn from a network of independent creatives and matched to the model category most relevant to their workflow, then asked to evaluate model outputs at each phase, shown under blinded labels (A/B/C/D) and in randomized order. Prompts chain across phases to mimic a real designer's process: Mockup and Refinement prompts build on the prior phase and supply a seed reference where applicable (an image for the image domains, the prior HTML for the code domains).
54
 
55
  | | |
56
  |---|---|
 
61
  | Models evaluated | 13 |
62
  | Evaluators | 31 (anonymized) |
63
  | Pairwise judgments | 3,174 |
64
+ | Scalar rating rows | 812 (code domains only — see note) |
65
+ | Qualitative feedback rows | 2,247 (2,116 per-output + 131 comparison-rationale — see note) |
66
  | Qualitative feedback rows | 2,247 |
67
 
68
  ### Domains, modalities, and models
 
81
  - **Mockup** — the chosen direction is realized: product shots, scene composition, brand identity.
82
  - **Refinement** — near production-ready; targeted edits for consistency and polish.
83
 
84
+ ## Identifiers and how to join
85
+
86
+ Opaque identifiers are released as deterministic UUIDs; the human-readable `prompt_id` slug (e.g. `ad_image_idea_4152`) is kept so domain/stage/sequence stay legible, and `domain`, `stage`, and `model_id` are readable throughout.
87
 
88
+ - **`prompt_id`** (readable) joins every table back to `prompts_workflow.csv`.
89
+ - **`content_id`** (UUID) is the per-asset primary key in `model_outputs.csv`. Pairwise rows reference assets directly via `content_id_left` / `content_id_right` / `content_id_chosen`, so you can identify the exact two outputs in any comparison without a join.
90
+ - **`evaluator_id`** (UUID) identifies an evaluator consistently across `pairwise_comparisons.csv`, `scalar_feedback.csv`, and `qualitative_feedback.csv` (31 evaluators, 1:1).
91
+ - **`evaluation_id`** (UUID) is shared by the scalar and qualitative tables. Scalar rows join 1:1 to the qualitative rows where `record_type = "per_output"`.
92
 
93
  ```python
94
  from datasets import load_dataset
95
 
96
+ scalar = load_dataset("contra-labs/HCB", "scalar_feedback", split="train").to_pandas()
97
+ qual = load_dataset("contra-labs/HCB", "qualitative_feedback", split="train").to_pandas()
98
 
99
+ # scalar <-> qualitative join is 1:1 on the per-output rows
100
+ per_output = qual[qual.record_type == "per_output"]
101
+ merged = scalar.merge(per_output, on="evaluation_id", suffixes=("_scalar", "_qual"))
102
  ```
103
 
104
+ ## Format and schema
105
+
106
+ Five CSV files, joinable as described above. `domain` and `stage` can be recovered for any record from `prompts_workflow.csv` or `model_outputs.csv`.
107
  ### `prompts_workflow.csv` — 95 rows (one per prompt)
108
 
109
+ | `prompt_id` | Readable prompt identifier (e.g. `ad_image_idea_4152`). |
110
+ | `raw_prompt_id` | Original source numeric id for the prompt. |
 
111
  | `domain` | One of the five creative domains. |
112
  | `stage` | `Ideation`, `Mockup`, or `Refinement`. |
113
  | `prompt_text` | The text prompt presented to the models. |
114
+ | `input_image` | Seed/reference image(s) for the image domains (Ad Images, Brand Design), as `https://media.contra.com/image/upload/<id>` URLs. Populated for 44 of 95 rows (mainly Mockup/Refinement); empty for most Ideation prompts. A few rows carry multiple seeds as a comma-and-space–delimited (`, `) list of URLs in one cell. |
115
+ | `input_html` | Seed/reference HTML for the code domains (Desktop App, Landing Pages) — the full prior-phase document the model builds on. Populated for the 26 code-domain Mockup/Refinement prompts; empty otherwise. |
116
 
117
  ### `model_outputs.csv` — 380 rows (one per prompt × model)
118
 
119
+ | `content_id` | Per-asset primary key (UUID). |
 
 
120
  | `prompt_id` | Foreign key to `prompts_workflow.csv`. |
121
  | `domain`, `stage` | Creative domain and workflow phase. |
122
  | `model_id` | Model that produced the output. |
123
+ | `model_output` | The generated artifact. **Format depends on modality:** image domains (Ad Images, Brand Design) store a `https://media.contra.com/image/upload/<id>` URL; Ad Video stores a `https://media.contra.com/video/upload/<id>` URL; code domains (Desktop App, Landing Pages) store the raw generated HTML inline. |
 
124
 
125
  ### `pairwise_comparisons.csv` — 3,174 rows
126
 
 
128
 
129
  | Column | Description |
130
  |---|---|
131
+ | `comparison_id` | Unique comparison identifier (UUID). |
132
  | `prompt_id` | Foreign key to `prompts_workflow.csv`. |
133
  | `domain`, `stage` | Creative domain and workflow phase. |
134
+ | `evaluator_id` | Anonymized evaluator identifier (UUID). |
135
  | `user_core_skill` | Evaluator's professional specialty (e.g. Brand Designer, Product Designer, Digital Marketer). |
136
  | `model_left`, `model_right` | The two models compared. |
137
  | `chosen_model` | The preferred model (always equal to `model_left` or `model_right`). |
138
+ | `content_id_left`, `content_id_right`, `content_id_chosen` | The specific `model_outputs.content_id` assets shown and selected. |
139
 
140
+ **Sampling note.** Each domain has four candidate models (six possible pairs). Left/right position bias is negligible (49.8% left / 50.2% right).
141
+ ### `scalar_feedback.csv` — 812 rows
142
 
143
+ 1-5 Likert ratings on three dimensions, ordered from most objective to most taste-driven. **Note:** scalar ratings were collected only for the two code domains (Desktop App, Landing Pages) in this release; image and video domains are not rated here.
144
 
145
  | Column | Description |
146
  |---|---|
147
+ | `evaluation_id` | Shared identifier with `qualitative_feedback.csv` (UUID). |
148
+ | `prompt_id`, `evaluator_id`, `user_core_skill`, `model_id` | Join keys and evaluator/model identifiers. |
 
 
 
149
 
150
  ### `qualitative_feedback.csv` — 2,247 rows
151
 
152
+ Open-ended written rationale, machine-coded into themes via a deductive coding pass (GPT-4o against a predefined codebook) that assigns themes, per-theme sentiment, and representative quotes.
153
 
154
  | Column | Description |
155
  |---|---|
156
+ | `evaluation_id` | Shared identifier with `scalar_feedback.csv` (UUID). Unique within `record_type = "per_output"`. |
157
+ | `record_type` | `per_output` (2,116 rows; one evaluator's rationale for a single output) or `comparison_rationale` (131 rows; one evaluator's overall rationale for a comparison, not tied to a single output). |
158
+ | `prompt_id`, `evaluator_id`, `user_core_skill`, `model_id` | Join keys and evaluator/model identifiers. |
159
+ | `model_label` | Blinded label shown to the evaluator (A/B/C/D) for `per_output` rows; empty for `comparison_rationale` rows. |
160
+ | `feedback_section` | The survey section the response answers (`Outcome` or `Usability`). |
161
+ | `feedback_question` | The exact prompt/question shown to the evaluator (previously embedded as `[Section \| question]` scaffolding inside the free text). |
162
+ | `raw_feedback` | The evaluator's free-text answer (scaffolding removed). For `comparison_rationale` rows the same text is denormalized across the comparison's candidate rows — filter on `record_type` to avoid double-counting in row-level analysis. |
163
+ | `assigned_themes` | JSON list of coded themes (e.g. `prompt_adherence`, `visual_hierarchy`, `typography`); keys match `theme_sentiment` and `key_quotes`. |
164
+ | `theme_sentiment` | JSON object mapping each theme to `positive`, `negative`, `neutral`, or `mixed`. |
165
  | `key_quotes` | JSON object mapping each theme to a supporting quote from the feedback. |
166
 
167
  ## Curation