Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- intent-classification
|
| 5 |
+
- lead-mining
|
| 6 |
+
- distillation
|
| 7 |
+
- embeddinggemma
|
| 8 |
+
- cpu
|
| 9 |
+
metrics:
|
| 10 |
+
- precision
|
| 11 |
+
- accuracy
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# Tweet Buy-Intent Head (embeddinggemma-2 + MLP)
|
| 15 |
+
|
| 16 |
+
A 199 KB classification head distilled from a Gemini 2.5 Flash teacher:
|
| 17 |
+
scores X/Twitter posts 0-1 for **buying intent** ("is there an app that…",
|
| 18 |
+
"anyone recommend…", "I wish there was…").
|
| 19 |
+
|
| 20 |
+
Base embedder (frozen): **google/embeddinggemma-2** (768-d, Apache-2.0).
|
| 21 |
+
The head itself is an MLP 768→64→1 (49k params) — runs on CPU.
|
| 22 |
+
|
| 23 |
+
## Results (held-out test, n=1,179)
|
| 24 |
+
|
| 25 |
+
| metric | score |
|
| 26 |
+
| --- | --- |
|
| 27 |
+
| precision@5 | 1.00 |
|
| 28 |
+
| precision@10 | 1.00 |
|
| 29 |
+
| accuracy @0.5 | 0.877 |
|
| 30 |
+
|
| 31 |
+
⚠️ Labels are Gemini 2.5 Flash-generated (no human gold set yet): the scores
|
| 32 |
+
measure agreement with the teacher (~90% human agreement per a 50-sample
|
| 33 |
+
audit), not ground truth. Complaint-based intent is under-detected (safe
|
| 34 |
+
direction for lead queues: misses recall, not purity).
|
| 35 |
+
|
| 36 |
+
## Usage
|
| 37 |
+
|
| 38 |
+
```python
|
| 39 |
+
import torch
|
| 40 |
+
from huggingface_hub import hf_hub_download
|
| 41 |
+
from sentence_transformers import SentenceTransformer
|
| 42 |
+
|
| 43 |
+
embedder = SentenceTransformer("google/embeddinggemma-2", device="cpu",
|
| 44 |
+
model_kwargs={"dtype": torch.float32})
|
| 45 |
+
head_path = hf_hub_download("flashclub/tweet-buy-intent", "head.pt")
|
| 46 |
+
head = torch.nn.Sequential(
|
| 47 |
+
torch.nn.Linear(768, 64), torch.nn.ReLU(), torch.nn.Dropout(0.2), torch.nn.Linear(64, 1))
|
| 48 |
+
head.load_state_dict(torch.load(head_path, map_location="cpu"))
|
| 49 |
+
head.eval()
|
| 50 |
+
|
| 51 |
+
with torch.no_grad():
|
| 52 |
+
emb = torch.tensor(embedder.encode(["is there an app for invoicing"], normalize_embeddings=True))
|
| 53 |
+
score = torch.sigmoid(head(emb)).item()
|
| 54 |
+
print(score) # 0.82
|
| 55 |
+
```
|
| 56 |
+
|
| 57 |
+
Trained on 7,866 X posts (2,850 intent) distilled from a Gemini teacher.
|
| 58 |
+
Pipeline: github.com/flashclub/x-lead-intent-pipeline
|