hikewa/eisv-lumen-teacher

EISV-Lumen Teacher — LoRA on Qwen3-4B

A LoRA fine-tune of Qwen/Qwen3-4B that maps EISV trajectory shapes to primitive expression tokens. Given a sequence of EISV dimension values and a classified trajectory shape, the model generates a structured expression — a short sequence of emotion-like tokens that a downstream embodied system (Lumen, running on a Raspberry Pi 4) uses to drive LED color, screen drawing, and inner-voice narration.

This is the teacher model in a teacher-student pipeline. Its outputs are distilled into a lightweight RandomForest student that runs on-device.

Model Details

Detail	Value
Base model	Qwen/Qwen3-4B
Method	LoRA (Low-Rank Adaptation)
Rank (r)	16
Alpha	32
Dropout	0.1
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Task type	CAUSAL_LM
Training examples	2,880 (50% real + 50% synthetic)
Training steps	1,260
Adapter size	126 MB
PEFT version	0.18.1
License	Apache 2.0

What is EISV?

EISV is a four-dimensional state space computed from real sensor readings on a Raspberry Pi 4 (BrainCraft HAT with BME280, VEML7700, CPU telemetry):

Dimension	Name	What it captures	Source signals
E	Energy	Warmth and capacity	CPU temperature, ambient temperature, neural beta+gamma
I	Information Integrity	Clarity and coherence	Prediction accuracy, neural alpha, world light, sensor coverage
S	Entropy	Uncertainty and disorder	Humidity deviation, memory usage, missing sensors, pressure deviation
V	Void	Disengagement and absence	Inverse of memory, CPU, and disk availability

These are not simulated or synthetic dimensions — they are derived from physical sensors and computational telemetry on a real device, sampled every 30 seconds.

Trajectory Shapes

A trajectory is a time-window of EISV values. The shape classifier identifies 9 distinct patterns:

Shape	Description	Real-data frequency
settled_presence	Stable, grounded state with low entropy and void	47%
convergence	Dimensions moving toward alignment	41%
entropy_spike_recovery	Sharp entropy increase followed by return to baseline	5%
basin_transition_up	Shift from lower to higher energy basin	2%
basin_transition_down	Shift from higher to lower energy basin	2%
rising_entropy	Sustained increase in uncertainty	1.5%
falling_energy	Gradual energy decline	1.5%
void_rising	Increasing disengagement	0.3%
drift_dissonance	Conflicting dimensional movements	synthetic only

The distribution is heavily skewed toward settled_presence and convergence because a well-functioning Pi is usually stable. The training set uses 50/50 real+synthetic blending to ensure the model sees enough rare shapes.

Expression Vocabulary

The model generates structured expressions using 15 primitive tokens and 5 patterns.

Tokens (15)

~warmth~ ~curiosity~ ~resonance~ ~stillness~ ~boundary~ ~reaching~ ~reflection~ ~ripple~ ~deep_listening~ ~emergence~ ~questioning~ ~holding~ ~releasing~ ~threshold~ ~return~

Patterns (5)

Pattern	Structure	Example
SINGLE	One token	~stillness~
PAIR	Two tokens	~warmth~ ~resonance~
TRIPLE	Three tokens	~curiosity~ ~reaching~ ~emergence~
REPETITION	Repeated token (intensity)	~warmth~ ~warmth~ ~warmth~
QUESTION	Token with ? (uncertainty)	~threshold~?

Evaluation Results

Evaluated on 500 real trajectories (no synthetic data in the test set):

Metric	Score
Mean coherence	0.952
Valid rate	100%
Pattern accuracy	25.8%

Coherence measures how well the generated expression fits the trajectory shape (semantic alignment between tokens and shape meaning). Valid rate measures whether outputs parse as legal expressions. Pattern accuracy measures exact match of the structural pattern — the relatively low score is expected since multiple patterns can be coherent for a given shape.

Per-Shape Coherence

Shape	Coherence
settled_presence	0.993
convergence	0.936
void_rising	1.000
basin_transition_down	1.000
basin_transition_up	1.000
rising_entropy	1.000
falling_energy	0.875
entropy_spike_recovery	0.833

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-4B",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

# Apply LoRA adapter
model = PeftModel.from_pretrained(base_model, "hikewa/eisv-lumen-teacher")

# Format input prompt
prompt = """You are an EISV expression generator. Given trajectory dynamics, produce a primitive expression.

Trajectory shape: settled_presence
EISV values: E=0.72, I=0.85, S=0.15, V=0.08
Dominant dimension: I (Information Integrity)
Trend: stable

Expression:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=32,
    temperature=0.7,
    do_sample=True,
)
result = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(result)
# Example output: ~stillness~ ~resonance~

Related

• Dataset: hikewa/unitares-eisv-trajectories — 2,880 trajectory-expression pairs used for training
• Student model: hikewa/eisv-lumen-student — distilled RandomForest that runs on Raspberry Pi 4
• Demo: hikewa/eisv-lumen-explorer — interactive Space to explore trajectories and expressions

Citation

@misc{eisv-lumen-teacher-2026,
  title={EISV-Lumen Teacher: LoRA Fine-Tune for Trajectory-to-Expression Generation},
  author={hikewa},
  year={2026},
  url={https://huggingface.co/hikewa/eisv-lumen-teacher},
  note={LoRA adapter on Qwen/Qwen3-4B mapping EISV trajectory shapes to primitive expression tokens}
}