README.md

metadata

title: README
emoji: 💻
colorFrom: pink
colorTo: red
sdk: static
pinned: false
license: mit
short_description: 'ISAAC OS — Neural v1 (Agentic-Lite deterministic evaluation '

ISAAC OS — Neural v1 (Deterministic Evaluation, Agentic-Lite)

Organization: Isaac-AI-OS
Model ID: isaac-20b
Policy Version: agentic-lite-v1
Artifacts Dataset: isaac-20b-eval-artifacts
Docker Digest: isaac-hf@sha256:6fc9f0d85dfe56daba8fc92496718226f056014b3e84ee7a823df1d9271a57c0

---

ISAAC is a self-verifying neural operating system designed for reproducible, auditable AI.
The Agentic-Lite evaluation mode enforces deterministic sampling (temperature=0, top_p=0, seed=7) and code-only normalization, producing byte-identical artifacts across runs.

🔍 Current Subset Results

Benchmark	Split	Metric	Score
HumanEval	N=5	pass@1	0.60
MBPP	N=5	pass@1	0.80
SWE-Bench Lite	1/1 instance resolved	resolved via fallback_dataset_patch	1 / 1

---

🧩 Reproducibility

• Deterministic “Agentic-Lite” mode: single plan, no concurrency, fixed seeds.
• Evaluation artifacts (LM, Code, SWE) are all published for cross-verification.
• Manifest pinned to Docker digest above for full audit trace.

📂 Artifacts

• LM normalized results
• Code benchmarks
• SWE-Bench Lite results
• Manifest

💡 Roadmap

• Add logprobs to /v1/completions → full MC reasoning (MMLU/ARC/HellaSwag).
• Enable kernel-level determinism for multi-node HA.
• Publish “Agentic Swarm” uplift appendix once replay bundles are live.