README.md

metadata

license: mit
language:
  - en
tags:
  - text-classification
  - multi-label-classification
  - intent-classification
  - healthcare
  - social-media
  - uk
library_name: transformers
pipeline_tag: text-classification

DocMap Leads Classifier (UK healthcare, social media)

Multi-head encoder classifier built on microsoft/deberta-v3-base to identify:

• Intent (single label)
• Symptoms (multi-label)
• Specialties (multi-label)

Trained on DocMap leads_v1 JSONL (Supabase), using simple weak labels derived from keywords/zero-shot prompts. See label_config.json for the canonical label spaces.

What’s in this repo

• model.safetensors: classifier heads + encoder weights
• label_config.json: lists of intents, symptoms, specialties
• tokenizer.json, tokenizer_config.json, special_tokens_map.json, spm.model
• README.md, .gitattributes
• (Checkpoints may be removed for smaller repo size)

Intended use

• Lead identification from public social media text in a UK healthcare context.
• Outputs: intent label and multi-label sets for symptoms/specialties.

Not a medical device. Do not use for diagnosis; for triage or marketing pre-filtering only.

Training

• Base: microsoft/deberta-v3-base
• Epochs: 3, batch size: 16, lr: 2e-5, max_len: 256
• Split: 10% validation
• Threshold sweep on validation suggested 0.3 as default for multi-label heads.

Quick start (Python)

This model uses a lightweight custom head. Load with the snippet below (no HF widget).

import json, torch
import torch.nn as nn
from transformers import AutoTokenizer, AutoModel

repo_id = "YOUR_USER_OR_ORG/docmap-leads-classifier-v1"  # change to your repo
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load labels
from huggingface_hub import hf_hub_download
cfg_path = hf_hub_download(repo_id, "label_config.json")
with open(cfg_path, "r") as f:
    cfg = json.load(f)

tokenizer = AutoTokenizer.from_pretrained(repo_id)

class LeadsClassifier(nn.Module):
    def __init__(self, base_model_name, num_intents, num_symptoms, num_specialties):
        super().__init__()
        self.encoder = AutoModel.from_pretrained(base_model_name)
        hidden = self.encoder.config.hidden_size
        self.dropout = nn.Dropout(0.1)
        self.intent_head = nn.Linear(hidden, num_intents)
        self.sym_head = nn.Linear(hidden, num_symptoms)
        self.spec_head = nn.Linear(hidden, num_specialties)
    def forward(self, input_ids=None, attention_mask=None):
        out = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        cls = self.dropout(out.last_hidden_state[:, 0, :])
        return {
            "intent_logits": self.intent_head(cls),
            "sym_logits": self.sym_head(cls),
            "spec_logits": self.spec_head(cls),
        }

model = LeadsClassifier(
    base_model_name="microsoft/deberta-v3-base",
    num_intents=len(cfg["intents"]),
    num_symptoms=len(cfg["symptoms"]),
    num_specialties=len(cfg["specialties"]),
).to(device)

# Load weights
sd_path = hf_hub_download(repo_id, "model.safetensors")
model.load_state_dict(torch.load(sd_path, map_location=device, weights_only=True))
model.eval()

def predict(texts, thr=0.3):
    batch = tokenizer(texts, padding=True, truncation=True, max_length=256, return_tensors="pt").to(device)
    with torch.no_grad():
        out = model(**batch)
        intent = torch.softmax(out["intent_logits"], dim=-1).argmax(dim=-1).tolist()
        sym_prob = torch.sigmoid(out["sym_logits"])
        spec_prob = torch.sigmoid(out["spec_logits"])
    intents = [cfg["intents"][i] for i in intent]
    symptoms = [[cfg["symptoms"][j] for j, p in enumerate(row) if p >= thr] for row in sym_prob.tolist()]
    specialties = [[cfg["specialties"][j] for j, p in enumerate(row) if p >= thr] for row in spec_prob.tolist()]
    return [{"intent": i, "symptoms": s, "specialties": sp} for i, s, sp in zip(intents, symptoms, specialties)]

print(predict(["Any advice on fever? Based in Glasgow, started 3 days ago."], thr=0.3))

Inference thresholds

• Default multi-label threshold: 0.3 (from validation sweep).
• Tune per use-case; 0.5 is stricter, 0.2 more sensitive.

Limitations and risks

• Weakly supervised labels; potential label noise and leakage.
• Social media domain; may not generalize to clinical text.
• Not for medical diagnosis or emergency advice.

License

• MIT (inherits base model’s MIT license).

Citation

Please cite the base model and this repository if you use it in research or production.