eisv-lumen-student

EISV-Lumen Student -- Distilled RandomForest for Edge Deployment

A lightweight RandomForest ensemble distilled from the EISV-Lumen Teacher (LoRA on Qwen3-4B). Achieves 0.986 coherence on the EISV expression-generation task while fitting in ~13 MB of JSON with zero external dependencies -- only Python stdlib required. Designed to run on a Raspberry Pi 4 (Lumen's physical host).

Model Details

How It Works

The student decomposes EISV expression generation into three chained classification problems, each solved by an independent RandomForest:

1. Pattern classifier -- predicts one of 5 expression patterns: SINGLE, PAIR, REPETITION, QUESTION, TRIPLE
2. Token-1 classifier -- predicts the primary EISV token from 15 classes (e.g., ~stillness~, ~warmth~, ~emergence~)
3. Token-2 classifier -- predicts the secondary token from 15 + none, conditioned on the Token-1 prediction (token1 index appended as extra feature)

The pattern determines how tokens are assembled into the final expression string (e.g., PAIR yields two distinct tokens, REPETITION repeats token-1 twice).

Results

Why does the student exceed the teacher? The RandomForest decision boundaries naturally cluster predictions toward high-affinity tokens for each trajectory shape. While the student disagrees with the teacher on exact token choices ~30% of the time, the tokens it picks are still coherent -- they belong to the same affinity region of EISV space. The coherence metric rewards any valid expression, not exact match.

The two coherence figures also come from different test sets: the student's from 1,080 held-out teacher-labelled synthetic windows, the teacher's from 500 real Lumen windows. Read the comparison as indicative, not as a like-for-like result.

Zero-Dependency Usage (recommended for edge)

The exported/ directory contains JSON-serialized forests and a standalone inference module. No pip packages required.

from student_inference import StudentInference

student = StudentInference("path/to/exported/")

result = student.predict("settled_presence", {
    "mean_E": 0.7, "mean_I": 0.6, "mean_S": 0.2, "mean_V": 0.05,
    "dE": 0.0, "dI": 0.0, "dS": 0.0, "dV": 0.0,
    "d2E": 0.0, "d2I": 0.0, "d2S": 0.0, "d2V": 0.0,
})
# result = {"pattern": "SINGLE", "eisv_tokens": ["~stillness~"],
#           "token_1": "~stillness~", "token_2": "none"}

Only json and os from the standard library are used. The inference module walks each decision tree node-by-node and averages class probabilities across all trees -- identical to sklearn's predict logic.

File Structure

eisv-lumen-student/
|-- README.md                  # This file
|-- training_metrics.json      # Cross-validation metrics
|-- eval_results.json          # Full evaluation results
|-- exported/                  # Zero-dependency JSON format
    |-- pattern_forest.json    # Decision trees as JSON (3.0 MB)
    |-- token1_forest.json     # Decision trees as JSON (4.5 MB)
    |-- token2_forest.json     # Decision trees as JSON (5.1 MB)
    |-- scaler.json            # Scaler parameters (511 B)
    |-- mappings.json          # Label mappings (1.1 KB)
    |-- student_inference.py   # Standalone inference (4.9 KB)

Training Details

• Distillation source: Teacher (Qwen3-4B LoRA v6, 0.952 coherence on real Lumen trajectories)
• Data generation: 4,320 synthetic EISV trajectories labeled by teacher inference (480 per shape x 9 shapes), plus 1,080 held-out test examples
• Forest hyperparameters: n_estimators=100, max_depth=None, random_state=42 (sklearn defaults)
• Feature engineering: 12 numeric features (4 EISV means + 4 first derivatives + 4 second derivatives) standardized via StandardScaler, plus 9-dimensional one-hot encoding of trajectory shape

Related

• Teacher model: hikewa/eisv-lumen-teacher -- Qwen3-4B LoRA that generated the training labels
• Dataset: hikewa/unitares-eisv-trajectories -- EISV trajectory data from Lumen
• Explorer Space: hikewa/eisv-lumen-explorer -- interactive demo

Citation

@misc{eisv-lumen-student-2026,
  title   = {EISV-Lumen Student: Distilled RandomForest for Edge Deployment},
  author  = {Wang, Kenny},
  year    = {2026},
  url     = {https://huggingface.co/hikewa/eisv-lumen-student},
  note    = {Knowledge-distilled RandomForest ensemble for EISV expression
             generation on Raspberry Pi}
}