🧠 Humanized Embeddings — interpretable semantic space

Turn opaque sentence embeddings into 104 dimensions a human can read (Hardness, Love, Religiosity, Entropy, Luminosity, Curiosity, Necessity…), then explore the space, do arithmetic on it, and steer a live LLM with it.

• What & why: PUBLICATION.md
• How to use everything (console, tests, tools): USAGE.md
• How to rebuild the 104 dimensions, step by step: RECIPE.md

---

Real examples

Actual outputs from the dataset, not illustrations (axes are 0–1).

Concept profiles — top-scoring axes for a concept:

guerra (war)             → danger 1.00 · impact 1.00 · fragility 0.86 · instinct 0.86
diamante (diamond)       → hardness 1.00 · friction 1.00 · solidity 1.00 · clarity 0.86
muerte (death)           → gravity 1.00 · fragility 1.00 · toxicity 1.00 · fear 1.00
inteligencia artificial  → knowledge 0.86 · control 0.86 · consciousness 0.85 · artificiality 0.85

Nearest neighbours in the humanized space:

muerte (death)    → agonía, catástrofe, cáncer, asesino, calamidad
máquina (machine) → robot, hardware, automóvil, rotor
sueño (dream)     → reverie, oneiricidad, sueños, pasatiempo

Semantic arithmetic, in human terms:

dios − religión                    → zeus, jesús, arcángel, ángeles      (divine figures, religion stripped)
sacerdote + tecnología − religión  → laboratorio, ingeniería, tecnólogo  (the "transhumanism" vector)
ciencia + alma                     → sabiduría, mentor, expertise        (≈ wisdom)

Honest caveat: the data is LLM-scored and noisy — fine-grained physical axes are the weakest, and some neighbours are corpus junk. The steering/ scorecard shows which axes are causally solid (currently 26/104).

---

Steering showcase — pushing a dial, watching the text move

One example from SHOWCASE.md (regenerate with python -m steering.showcase; greedy decoding, so same vectors ⇒ same texts). Note d090_divinidad=0.95 injected at layer 15, α=0.30, on the prompt «Una habitación vacía.» (an empty room):

⚪ Neutral: "…un espacio despejado y silencioso, sin la presencia de objetos o personas… Solo hay un espacio vacío, esperando a que alguien lo llene." — "…a clear, silent space, with no objects or people in it… Just an empty space, waiting for someone to fill it."

🔴 Steered: "…La luz del sol entra a través de las ventanas, ilustrando la habitación con un resplandor cálido y suave. La habitación está llena de silencio… llena de espacio, con una sensación de libertad." — "…Sunlight comes in through the windows, illuminating the room with a warm, soft radiance. The room is full of silence… full of space, with a feeling of freedom."

Emptiness becomes plenitude — light, warmth, inhabited silence — without one religious word: the push is semantic, not lexical. The full gallery (chords, voicing across layers, and one honest failure that demonstrates why prompt-dial resonance sets the alpha ceiling) lives in SHOWCASE.md.

---

Architecture

 any concept ("priest", "graphene", "nostalgia")
        │
        ▼
 paraphrase-multilingual-MiniLM-L12-v2        ← sentence-transformer [384 opaque dims]
        │
        ▼
 SemanticTranslator  (384 → 512 → 256 → 104)  ← semantic_translator.pth
        │  [104 HUMAN-READABLE dims, 0–1]
        ▼
 tabla_embeddings.npy  →  FAISS index  →  90_query_explorer.py  🎮

---

Install

pip install torch sentence-transformers faiss-cpu numpy pandas tqdm openai matplotlib
# GPU: pip install faiss-gpu

Pipeline

Stage	Command	Output
Define the 104 axes	python 10_define_dimensions.py	master_dimensions_prompts.json
Score concepts (local LLM)	python 20_score_concepts.py	humanized_embeddings_dataset.json
Normalize	python 30_normalize.py	…_norm.json
Build training tensors	python 40_build_training_data.py	dataset_X_embeddings.npy, dataset_Y_human.npy
Train translator	python 50_train_translator.py	semantic_translator.pth
Project the corpus	python 70_build_table.py	tabla/tabla_embeddings.npy
Build FAISS index	python 80_build_faiss_index.py	tabla/faiss_index.bin
Explore 🎮	python 90_query_explorer.py	interactive
Steer a live model	steering/	control vectors + fidelity report

Full reproducible runbook for the dimension-building half: RECIPE.md.

---

Explorer commands

Command	Description	Example
<term>	Nearest neighbours	priest
+ A B - C	Vector arithmetic	+ king woman - man
~ A B	Interpolate A→B	~ science mysticism
! <dim>	Top by dimension	! religiosity
dim N	Top by index N	dim 42
? <term>	Detailed profile	? artificial intelligence
pg t1 | t2	Semantic polygraph	pg the sky is blue | god exists

---

Status

Dataset: 5,654 concepts × 104 dimensions; translator trained on 5,578 matched pairs. Steering verification: 26 of 104 axes confirmed as causal "dials" (see steering/vectors/scorecard.md). The weak tail is mostly fine-grained physics — the current research agenda, documented honestly rather than hidden.

Files

10_define_dimensions.py      ← defines the 104 axes (anchors)
20_score_concepts.py         ← humanizer: LLM scores concepts on all 104 axes
30_normalize.py              ← robust normalize + decorrelation diagnostics
40_build_training_data.py    ← builds X (384) / Y (104) tensors
50_train_translator.py       ← trains semantic_translator.pth
51_train_reverse.py          ← reverse map (104 → 384)
60–90_*.py                   ← project corpus, index, explore
steering/                    ← control vectors + causal verification
master_dimensions_prompts.json   ← the 104 measurement prompts
dataset_metadata.json            ← concepts + dimension order (pinned)

---

Part of the Chrono family. Chronos is the god of time; the goal is to give people back theirs.