mjpsm/Bisa-xgb-model

Bisa Soulprint Regression Model

Model Details

• Model Name: Bisa Soulprint Regression (XGBoost)
• Author: Mazamesso “Mazo” Meba / Coding in Color (MyVillage Project)
• Framework: XGBoost Regressor + SentenceTransformer embeddings
• Embedding Model: all-mpnet-base-v2
• Output: Continuous score between 0.01–0.99

---

Intended Use

This model is part of the Soulprint Archetype system, which generates continuous scores (0.01–0.99) for the Bisa archetype. Bisa relates to endurance, resilience, and the capacity to persist within community and cultural contexts.

The model is trained on 1,308 balanced examples, evenly distributed across low (0.01–0.39), mid (0.40–0.69), and high (0.70–0.99) ratings. Inputs are short text statements (1–5 sentences), and outputs are continuous values representing Bisa intensity.

Example Applications

• Cultural storytelling: Scoring responses by levels of resilience and persistence.
• Community dialogue analysis: Identifying strength of endurance themes.
• Educational tools: Teaching Afrocentric archetype values in AI-powered environments.

---

Training Data

• Dataset Size: 1,308 rows
• Balance:
  • Low: 436
  • Mid: 436
  • High: 436
• Source: Synthetic, culturally-informed dataset generated using prompt engineering and controlled vocabulary/persona variation.

Labels

• Low (0.01–0.39): silence, avoidance, withholding
• Mid (0.40–0.69): mild questioning, observations, suggestions
• High (0.70–0.99): confrontation, criticism, persistence, challenge

---

Training Procedure

• Embeddings: all-mpnet-base-v2 SentenceTransformer
• Regressor: XGBoost
  • n_estimators=1200
  • learning_rate=0.03
  • max_depth=10
  • subsample=0.9
  • colsample_bytree=0.9
  • reg_alpha=0.5
  • reg_lambda=5

---

Evaluation

• Test Size: 20% split (stratified by label bins)
• Metrics:
  • MSE: 0.0202
  • R²: 0.733

This means the model explains ~73% of the variance in Bisa scores.

Comparative Context

• Other Soulprint archetype models (Nzinga, Kinara, Ubuntu) achieved R² in the 0.80–0.82 range.
• Bisa currently performs at R² = 0.73, making it solid but with room for improvement via hyperparameter tuning or additional training data.

---

Limitations

• Synthetic data bias: While diverse, the dataset is synthetic and may not cover the full spectrum of cultural narratives.
• Moderate performance: R² = 0.73 means ~27% of variance is unexplained. Model works, but may misclassify edge cases.
• Context dependency: Model is tuned for Soulprint archetype research and may not generalize well to unrelated domains.

---

Future Work

• Hyperparameter tuning (Optuna/GridSearch for deeper optimization).
• Expanded dataset (adding 500+ new examples, including edge anchors at 0.05 and 0.95).
• Ensemble testing (combining Bisa with other archetype regressors).

---

Example Usage

import xgboost as xgb
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download

# -----------------------------
# 1. Download model from Hugging Face Hub
# -----------------------------
REPO_ID = "mjpsm/Bisa-xgb-model"
FILENAME = "Bisa_xgb_model.json"

model_path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME)

# -----------------------------
# 2. Load model + embedder
# -----------------------------
model = xgb.XGBRegressor()
model.load_model(model_path)

embedder = SentenceTransformer("all-mpnet-base-v2")

# -----------------------------
# 3. Example prediction
# -----------------------------
text = "Despite noticing errors, I stayed quiet and accepted the outcome."
embedding = embedder.encode([text])
score = model.predict(embedding)[0]

print("Predicted Bisa Score:", round(float(score), 3))