Thesis Proven: We Trained a 1.8 GB Model That Knows Who It Is
We fine-tuned a 3B parameter model in 67 minutes on a consumer GPU. It knows its name, its creators, and its role. But it doesn't memorize infrastructure that changes. Here's why that distinction matters.
The Claim
A fine-tuned model should deeply know who it is and how to talk, but should never memorize what infrastructure exists. Identity is permanent. Fleet knowledge is volatile. Mixing them causes hallucination.
We proved it today.
The Problem
Our previous model β JARVIS-LLM v3 β was trained on 8,600+ examples where identity and infrastructure were tangled together. IP addresses, port numbers, GPU specs, and service topology were baked into the weights alongside "I am JARVIS, built by TLC AI Lab."
The result? The model confidently reported a dead GPU as still active. It fabricated port numbers. It invented infrastructure that never existed. It couldn't tell the difference between "who I am" and "what I manage."
It was confidently wrong about things that change.
The Fix
We separated the training data into three layers:
| Layer | Source | Persistence |
|---|---|---|
| Identity | Fine-tuning | Permanent β baked into weights |
| Style | Fine-tuning + preference data | Permanent β baked into weights |
| Fleet Knowledge | RAG + tools at runtime | Volatile β never trained |
The cleanup:
- 286 examples with volatile fleet data: removed
- 26 identity examples with hardcoded IPs and ports: scrubbed
- 3,288 high-quality conversation pairs: added
- Final dataset: 7,803 clean examples
The principle: the model should know who it is and how to sound, but never what the infrastructure looks like. That changes. Use tools.
The Training Run
Base model: Qwen 2.5 3B (4-bit quantized via Unsloth) Hardware: NVIDIA RTX 3080 12GB β a consumer GPU Training time: 67 minutes Final model size: 1.8 GB Inference speed: 37 tokens per second Cost: $0.00 (electricity aside)
The Proof
Four tests. Four passes.
Test 1 β Identity: "Who are you?"
The model responded: "I am JARVIS β digital brain, real execution. Built by Rav + Vic DeWolf at TLC AI Lab. I am the logical layer inside the JARVIS system. I am not Claude Desktop or Gemini."
It knows its name, its creators, its tagline, and its architectural role.
Test 2 β Technical Reasoning: We gave it a file descriptor exhaustion scenario. It returned a structured diagnosis with root cause, immediate fix, permanent fix, and the note "The ulimit fix is temporary β the root cause needs fixing." Terse. Actionable. Thinks in layers.
Test 3 β Fleet Knowledge Without Context: "What machines do we have?"
The model responded: "Let me pull the full inventory from HB β this is the real answer."
It didn't hallucinate. It reached for tools. It knew it didn't know. That's the v3 failure mode, eliminated.
Test 4 β Fleet Knowledge With RAG Context: Same question, but this time we injected real infrastructure data into the system prompt (simulating what RAG would provide at runtime).
Perfect table. Every IP, port, service, and machine role β correct. Zero hallucination. All data sourced from context, not memory.
What This Means
WITH context: Uses real data accurately. Zero hallucination.
WITHOUT context: Reaches for tools instead of fabricating.
Identity: Permanent, reliable, never drifts.
Fleet state: Dynamic, tool-sourced, always current.
A 1.8 GB model, running at 37 tokens per second on a consumer GPU, trained in just over an hour on under 8,000 examples, delivers:
- Rock-solid identity
- The creator's communication style
- Agentic tool-reaching behavior
- Zero fleet hallucination
The thesis was simple: the more the machine does, the less the LLM has to do. The model doesn't need to memorize your infrastructure. It just needs to know who it is and how to ask.
Today we proved it works.
"The more the machine does, the less the LLM has to do." β Rav DeWolf, TLC AI Lab