--- license: mit tags: - intent-classification - lead-mining - distillation - embeddinggemma - cpu metrics: - precision - accuracy --- # Tweet Buy-Intent Head (embeddinggemma-2 + MLP) A 199 KB classification head distilled from a Gemini 2.5 Flash teacher: scores X/Twitter posts 0-1 for **buying intent** ("is there an app that…", "anyone recommend…", "I wish there was…"). Base embedder (frozen): **google/embeddinggemma-2** (768-d, Apache-2.0). The head itself is an MLP 768→64→1 (49k params) — runs on CPU. ## Results (held-out test, n=1,179) | metric | score | | --- | --- | | precision@5 | 1.00 | | precision@10 | 1.00 | | accuracy @0.5 | 0.877 | ⚠️ Labels are Gemini 2.5 Flash-generated (no human gold set yet): the scores measure agreement with the teacher (~90% human agreement per a 50-sample audit), not ground truth. Complaint-based intent is under-detected (safe direction for lead queues: misses recall, not purity). ## Usage ```python import torch from huggingface_hub import hf_hub_download from sentence_transformers import SentenceTransformer embedder = SentenceTransformer("google/embeddinggemma-2", device="cpu", model_kwargs={"dtype": torch.float32}) head_path = hf_hub_download("flashcoder/tweet-buy-intent", "head.pt") head = torch.nn.Sequential( torch.nn.Linear(768, 64), torch.nn.ReLU(), torch.nn.Dropout(0.2), torch.nn.Linear(64, 1)) head.load_state_dict(torch.load(head_path, map_location="cpu")) head.eval() with torch.no_grad(): emb = torch.tensor(embedder.encode(["is there an app for invoicing"], normalize_embeddings=True)) score = torch.sigmoid(head(emb)).item() print(score) # 0.82 ``` Trained on 7,866 X posts (2,850 intent) distilled from a Gemini teacher. Dataset: https://huggingface.co/datasets/flashcoder/tweet-buy-intent-leads Pipeline: github.com/flashclub/x-lead-intent-pipeline