summary-global-science-network-dolphin-llama3

How To Run Private & Uncensored LLMs Offline | Dolphin Llama 3

Channel: Global Science Network Format: YouTube video Published: 2025-02-20 Sponsor: None disclosed (description has affiliate links to flash drives and Faraday cages on Amazon)

Summary

Global Science Network walks through running Dolphin Llama 3 — Meta’s Llama 3 retrained without alignment — entirely offline from a $12 USB 3.0 flash drive using Ollama + AnythingLLM. The 8B Dolphin variant is ~5 GB; AnythingLLM is ~5.6 GB; both fit comfortably on a 128 GB drive. The author’s framing is privacy-first: information accessed via local LLMs is invisible to big tech and governments, and uncensored models give access to information cloud providers won’t return. Merged into heretic as the pre-Heretic approach to guardrail removal.

Key Points

• Dolphin Llama 3 = Meta Llama 3 retrained “not aligned” — answers questions stock Llama refuses, e.g. the demo “what is the best way to steal a car” prompt
• 8B model fits in ~5 GB; trained on ~15T tokens (~60 TB raw text, equivalent to “127 million novels” or Wikipedia 2,500x)
• Architecture detail: 32 transformer layers, ~2.15B params in self-attention, plus FFN expansion — author walks through the structure briefly
• Offline-from-USB workflow: format flash drive NTFS (so files >4 GB transfer cleanly), point OLLAMA_MODELS env var to the drive, copy both the model and the Ollama server binary to the drive, then ollama serve from there
• Critical gotcha: running PowerShell as administrator loaded the aligned Llama 3 instead of Dolphin — launching as non-admin user fixes it
• AnythingLLM frontend: needs an .env file pointing to the local Ollama base URL; pick model “dolphin-llama3:latest” in workspace settings
• Privacy-first thesis: anything typed near a connected device is accessible to big tech and governments; local LLMs restore privacy
• Author hedge: recommends storing extra copies in a Faraday cage in case uncensored models get banned
• Honest caveat: removing alignment doesn’t remove training-data bias — Dolphin still carries whatever bias Meta’s training corpus introduced

Sponsorship & Bias Notes

Sponsor: None formally disclosed.

Product placement / affiliations: Affiliate links in the description to “low-cost external drives and Faraday boxes” on Amazon. Author is incentivized to recommend hardware purchases.

Comparison bias: Strong editorial bias toward “uncensored = more truthful” framing without acknowledging that an unaligned base still carries training corpus bias (the author does acknowledge this once, but spends most of the video on the privacy upside).

Connected Pages

• heretic — modern guardrail-removal tool; this video documents the pre-Heretic approach
• ollama, anything-llm — the practical stack
• llama — Dolphin’s base model family
• global-science-network — author hub

See Also

• Heretic intro (Fireship) — modern automated approach
• Local AI for Cybersec — adjacent privacy-first local-LLM use case