slendermantvb/jarvis-v2.5.0

Jarvis v2.5.0

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

This release keeps the same stable neural core while improving the runtime around it with adaptive long-input selection, stricter factual fallback, RAG, symbolic/Lisp reasoning, code verification and Reflexion-style memory.

Architecture

• Parameters: approximately 79M
• Vocabulary: 16,000 tokens
• Native 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

Improvements in v2.5

• Adaptive long-input handling: long prompts are reduced by selecting relevant original excerpts while preserving the final instruction/question.
• Factual guard: in the full local application, factual questions use deterministic facts and RAG before the neural fallback; unsupported factual questions return an explicit unknown response instead of being guessed.
• RAG memory: the local application indexes more than 50,000 retrievable items, including curated examples and project documentation.
• Coherence guard: rejects clearly corrupt, repetitive, internal-tag or excessively long outputs.
• Verified programming: Python examples can be checked with a safe AST interpreter before being accepted.
• Reflexion + verified learning: mistakes and user corrections can be recalled; only verified code examples are saved into the verified-learning store.

Local validation

Jarvis v2.5.0 passed the current project release tests:

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

These are project-specific local benchmarks, not a claim of universal accuracy.

Usage

Install:

pip install torch tokenizers safetensors

Run:

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

The included inference.py already uses the adaptive context selector and coherence guard.

Remaining limitations

The raw neural core is still a compact ~79M-parameter model, so it can make mistakes on difficult or unfamiliar tasks. Its native context window remains 512 tokens; v2.5 mitigates this for long inputs by selecting relevant excerpts, but that is not the same as a natively trained long-context model.

The full RAG, symbolic reasoning, factual guard, Reflexion memory and code-verification stack belongs to the local Jarvis application; the minimal raw-checkpoint inference in this repository does not reproduce every application-level safeguard.

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.