slendermantvb/jarvis-v2.4.1

Jarvis v2.4.1

Jarvis is a small experimental local language model developed by slendermantvb.

This repository contains the neural checkpoint and the minimal code required to load it. The full local Jarvis application additionally uses machine learning routing, symbolic/Lisp reasoning, RAG, code verification and Reflexion-style memory; those runtime components are not required to load the base neural model.

Architecture

• Parameters: approximately 79M
• Vocabulary: 16,000 tokens
• Context window: 512 tokens
• Hidden size: 512
• Layers: 10
• Attention heads: 8
• KV latent dimension: 128
• Feed-forward size: 1,365
• Mixture of Experts: 4 experts, Top-2 routing
• Shared expert: enabled
• Multi-token prediction depth: 2
• Framework: PyTorch

The architecture uses a compact latent-KV attention block, RoPE, SwiGLU and a small MoE.

Local validation

Jarvis v2.4.1 passed the project's current local release tests:

• Programming suite: 540 / 540
• Anti-hallucination suite: 120 / 120

These are project-specific local benchmarks. They must not be interpreted as 100% accuracy on programming, factuality, reasoning, or general language tasks.

Usage

Install:

pip install torch tokenizers safetensors

Run:

python inference.py "Hola, ¿quién eres?"

Files

• model.safetensors — model weights
• config.json — neural architecture configuration
• tokenizer.json — tokenizer
• model.py — custom PyTorch model
• config.py — configuration dataclass
• inference.py — minimal greedy inference example

Limitations

Jarvis is a research/learning project and is a very small language model by modern standards. It can produce incorrect, incomplete or incoherent responses. The 512-token context window is also restrictive for long documents and large codebases.

The local Jarvis application uses deterministic verification and external retrieval to compensate for some of these limitations. Those safeguards are not automatically present when using only the raw neural checkpoint in this repository.

License

The project-authored code and model artifacts in this repository are released under the Apache License 2.0, to the extent the author has rights to license them.

Training-data provenance remains subject to third-party terms. The original pretraining corpus includes Wikimedia/Wikipedia text, whose source content is subject to licenses including CC BY-SA and, depending on the material, GFDL. Apache-2.0 does not replace or override those third-party obligations. See NOTICE for provenance notes.