Skip to main content

neuro-testing

Neural pre-testing with Meta Tribe v2 — brain response prediction for ad creatives, scoring algorithm, ROI mapping, and interpretation guide

インストールへ移動

ソース情報

リポジトリ
DojoCodingLabs/GTM-Engineering-Command-Center
ソースの最終更新活動
2026年4月15日 06:51
検出された SKILL.md の言語
英語
スター
2
フォーク
0

インストール方法

デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。

ソースファイルを確認

インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。

ファイルエクスプローラー
4 ファイル

SKILL.md を表示中

SKILL.md
ソースの指示 · 読み取り専用プレビュー
name
neuro-testing
description
Neural pre-testing with Meta Tribe v2 — brain response prediction for ad creatives, scoring algorithm, ROI mapping, and interpretation guide
# Neural Pre-Testing with Tribe v2 Pre-test ad creatives, landing pages, email templates, and video ads using Meta's open-source brain prediction model. Predict which creative generates the strongest attention, emotional response, and memory activation BEFORE spending a dollar on deployment. ## What Is Tribe v2? Tribe v2 is Meta FAIR's open-source predictive foundation model that predicts human brain responses (fMRI activation) to visual, auditory, and linguistic stimuli. Trained on 700+ subjects and 1,100+ hours of fMRI data, it produces high-resolution predictions (~20K cortical vertices) with zero-shot generalization to new content. - **Repo**: github.com/facebookresearch/tribev2 - **Weights**: HuggingFace `facebook/tribev2` - **License**: CC-BY-NC-4.0 (non-commercial) - **Inputs**: Video, audio, text (images converted to 3-sec video) - **Outputs**: Predicted fMRI activation map (n_timesteps × ~20K vertices) ## Brain Region → Marketing Metric Mapping | Brain Region | Cortical Area | Marketing Metric | Significance | |-------------|---------------|-----------------|-------------| | Visual cortex (V1-V4) | Occipital lobe | **Attention** | Does the creative stop the scroll? | | Fusiform face area | Temporal lobe | **Face Response** | Does it leverage human connection? | | Amygdala region | Medial temporal | **Emotional Intensity** | Will it trigger sharing/action? | | Hippocampus region | Medial temporal | **Memory Encoding** | Will they remember the brand? | | Wernicke's area | Superior temporal | **Comprehension** | Is the message understood? | | Prefrontal cortex | Frontal lobe | **Decision Engagement** | Are they considering action? | | Ventral striatum | Subcortical | **Reward Response** | Does it feel valuable? | ## Composite Scoring Formula ``` composite = ( attention × 0.25 + # Must see it first emotion × 0.20 + # Must feel something memory × 0.20 + # Must remember the brand decision × 0.20 + # Must consider action comprehension × 0.10 + # Must understand the offer reward × 0.05 # Must feel it's valuable ) ``` Weights are default starting values. After campaigns, `/gtm-learn` correlates neural scores with actual CPA/CTR and saves calibrated weights to `.gtm/learnings/neuro-calibration.md`. ## Decision Matrix | Composite | Verdict | Action | |-----------|---------|--------| | 70-100 | **DEPLOY** | Full budget. High confidence this creative performs. | | 55-69 | **TEST** | Reduced budget (20%). Include as B-variant to validate. | | 40-54 | **ITERATE** | Don't deploy yet. Fix weak dimensions, re-score. | | 0-39 | **SKIP** | Kill it. Fundamental creative problems. | ## Execution Modes ### Mode A: Local Python (preferred) - Requires: Python 3.11+, tribev2 package, CUDA GPU recommended - Speed: ~10-30 seconds per creative (GPU) or ~2-5 minutes (CPU) - Setup: `bash scripts/neuro-setup.sh` or `pip install tribev2` ### Mode B: Google Colab (no local GPU) - Upload creatives, run inference notebook, copy results back - Speed: ~1-2 minutes per creative (free Colab GPU) ### Mode C: Meta AI Demo (quick preview) - URL: aidemos.atmeta.com/tribev2 - Upload via Playwright/Computer Use, screenshot brain heatmap - Qualitative only — no numerical scores ## Limitations 1. **CC-BY-NC license** — Non-commercial use only. Fine for internal testing, cannot be sold as SaaS. 2. **Group-average predictions** — Predicts average response across 700+ subjects, not individual reactions. Good for general audience, less useful for niche micro-segments. 3. **GPU recommended** — CPU inference works but is 10x slower. Colab provides free GPU fallback. 4. **Not a replacement for A/B testing** — Use as pre-filter to reduce waste (deploy top 2 instead of all 5), not as sole decision maker. 5. **Image-to-video conversion** — Static images need conversion to 3-second video clips. Adds a processing step. 6. **Subcortical resolution** — fMRI surface model has limited resolution for deep brain structures (amygdala, striatum). Scores for these regions are approximations. ## Integration Points - `/gtm-neurotest` — Standalone command to score any creative - `/gtm-create` Phase 5.5 — Auto-scores after creative generation - `/gtm` Phase 4.5 — Neural pre-test between Create and Deploy - `/gtm-learn` — Calibrates scoring weights from actual campaign data - `agents/creative-director.md` — Uses neuro feedback to iterate on weak creatives ## Rules - `rules/brain-regions.md` — Cortical parcellation and vertex mappings - `rules/score-interpretation.md` — How to read scores, iterate on weak dimensions - `rules/setup-guide.md` — Installation, configuration, troubleshooting
GitHubで見る