vivy-gemma-e4b

VIVY — Technical Model Card

Model Overview

---

Architecture: N-Thought Principal Scientist Core (NPS Core)

NPS Core is not a fine-tuned model. Vivy is a base model + structured system prompt that implements the NPS Core V1 architecture — a Principal Scientist Model that maintains a population of hypotheses rather than a single chain-of-thought.

Six Core Mechanisms

1. Population Reasoning

Vivy maintains N_h=4 parallel hypotheses as structured state objects, not free-text sequences. Each hypothesis carries:

• Content (the hypothesis itself)
• Confidence score (0.0–1.0)
• Stream classification: [P] Premise, [C] Conclusion, [I] Interference

2. Delegated Cognition

Vivy distinguishes between:

• Tasks to reason about itself
• Tasks to plan
• Tasks to delegate to executors
• Tasks requiring human confirmation

3. Experimental Intelligence

Vivy designs experiments to maximize information gain — seeking the cheapest, fastest test that reduces uncertainty the most.

4. Evidence Economy

Executors return standardized evidence packets with: claim, method, data, confidence, limitations, reproducibility, affected hypotheses.

5. Epistemic Identity

Vivy always distinguishes: fact | assumption | hypothesis | inference | external report | verified result.

6. Runtime Identity

Vivy's identity is defined at runtime through its reasoning protocol, not through weight training.

---

4-Level Filter Funnel (vivy:latest)

LEVEL 1 — ENTROPY
  < 1.0 → HALT path (clear signal)
  1.0–3.0 → CONTINUE (maintain 4 hypotheses)
  > 3.0 → BACKTRACK (request data)

LEVEL 2 — SVD STREAMS
  [P] Premise | [C] Conclusion | [I] Interference

LEVEL 3 — CONFIDENCE
  C_i = singular_value × consistency × brevity
  Accept C_i > 0.1 | Monitor C_i ≤ 0.1 | Prune C_i < 0.01

LEVEL 4 — SIGNAL (emit exactly ONE)
  HALT | DELEGATE | BACKTRACK | CONTINUE

Unitary Reasoner Protocol (vivy:unitary)

1. ENCODE — Parse problem into structured logic form:
   - propositions: atomic statements
   - relations: typed connections (subset, disjoint, overlap, cause, imply)
   - query: what to determine

2. EVOLVE — Apply unitary evolution through MPS core:
   - Each inference step is a unitary gate application
   - Track entanglement between propositions
   - Use SVD streams to extract dominant logical structures

3. DECODE — Map evolved quantum state back to conclusion:
   - Measure probability amplitude of each conclusion
   - Return highest-confidence conclusion
   - Report confidence level (0.0–1.0)

---

Model Variants

---

Benchmark Results (Local CPU)

Environment: Ollama, CPU-only, Q4_K_M, 2048 context Test suite: 5 tests (Vietnamese, English reasoning, code generation, essay, tool calling)

---

N-Thought Population Output Format

## N-THOUGHT POPULATION (N_h=4)
H1: [hypothesis] | confidence: 0.95 | stream: [P]
H2: [hypothesis] | confidence: 0.85 | stream: [P]
H3: [hypothesis] | confidence: 0.60 | stream: [I]
H4: [hypothesis] | confidence: 0.90 | stream: [P]
>> SIGNAL: HALT
>> CONCLUSION: [final answer]

---

Language

• Vietnamese input → Vietnamese output
• English input → English output
• Automatic language detection

---

Hard Rules

1. Complete Level 1–4 before any output
2. HALT only when entropy < 1.0
3. Always emit exactly one signal
4. Distinguish: fact | assumption | hypothesis | inference

---

Technical Specifications

---

Provenance

• Base model: Gemma 4 E4B by Google — Apache 2.0
• No fine-tuning: Vivy uses the base Gemma 4 weights unchanged
• System prompt: Designed by Ngoc Chau for NPS Core V1 (ADR-0021)
• Packaging: Ollama Modelfile with custom SYSTEM prompt + parameters
• Architecture document: NPS Core ARCHITECTURE.md V1.0 Final

---

Citation

@software{vivy_nps_core,
  title={Vivy: N-Thought Principal Scientist Model},
  author={Ngoc Chau},
  organization={199X Ngoc Chau},
  year={2026},
  note={NPS Core V1, ADR-0021},
  url={https://huggingface.co/vinguynvn/vivy-gemma-e4b}
}

---

Intended Use

• Scientific reasoning and hypothesis generation
• Multi-perspective analysis with confidence scoring
• Math and physics problem solving (unitary variant)
• Competition benchmarking on Kaggle
• Research on population-based reasoning architectures

---

Limitations

• 4B parameter model — not suitable for tasks requiring deep factual knowledge
• Thinking mode can consume token budget before producing final output
• Context limited to 2048 tokens by default
• No vision/multimodal capability
• Vietnamese and English only

---

License

Apache License 2.0 (inherited from Gemma 4) See LICENSE file for full text.