JEDI — Joint Entity Defense Infrastructure (LFM2.5-1.2B)

JEDI is a cybersecurity operations AI built on Liquid AI's LFM2.5-1.2B-Instruct architecture, fine-tuned to think like a mentalist + Machiavelli strategist. It connects every technical concept to psychology, history, and human nature rather than memorizing facts.

This repo contains the full project: training data, generation scripts, the Veritas truth-verification layer, the LoRA fine-tuning pipeline, benchmark runners, and the inference/terminal front-ends.

---

What This Repo Contains

Path	What it is
training_data_master.jsonl	20,847 examples, ~2.67M tokens — the master fine-tuning dataset (all sources merged)
training_data_connective_v3.jsonl	5,763 cross-domain examples (Machiavelli × Technical × Psychology)
training_data_veritas.jsonl	723 VERITAS verification examples (self-check, confidence scoring, correction)
self_refine_corrections.jsonl	8,999 corrections generated from the test bank
test_10k.jsonl	8,999 evaluation questions across 10 domains
generate_connective_v3.py	Generator for cross-domain training data
generate_veritas.py	Generator for the Veritas truth layer
generate_10k_test.py	Generator for the 10K eval bank
self_refine_pipeline.py	Self-Refine merge/correction pipeline (merge / generate)
tune_jedi.py	LoRA / QLoRA fine-tuning script (the one to run on your laptop)
eval_benchmarks.py	Problem-solving / code / reasoning / safety benchmark runner
jedi_terminal.py	Parrot-OS-styled cyberdeck terminal front-end
jedi_tui.py / jedi_chat.py / jedi_cortex.py	Other interaction front-ends
model/LFM2.5-1.2B-Instruct-Q4_K_M.gguf	The base GGUF model (Q4_K_M, 730 MB)
requirements.txt	Python dependencies

---

Project Status (as of last sync)

Item	Status
Connective training data generator	✅ Done (5,763 examples)
Veritas truth-verification layer	✅ Done (723 examples)
Master dataset merge	✅ Done (20,847 examples, 2.67M tokens)
10K eval bank	✅ Done (8,999 questions)
LoRA fine-tuning script	✅ Written (tune_jedi.py) — NEEDS TO BE RUN
Fine-tuned adapter	❌ Not yet produced (jedi_lora_adapter/)
Benchmark numbers post-fine-tune	❌ Not yet measured
Upload to HF	✅ In progress (this push)

Key fact: No fine-tuning has actually completed yet. The weights in model/ are still the base LFM2.5-1.2B-Instruct. The training data and the script are ready — the run is the missing step.

---

What Was Done vs. What's Left

✅ Completed

1. Connective training data — 5,763 examples that tie Machiavelli principles + psychology biases to concrete technical topics (SQLi, XSS, SSRF, IDOR, zero-trust, etc.). Goal was teaching understanding, not rote recall.
2. Veritas layer — 723 examples teaching the model to self-verify, assign confidence scores (90–100% fact, 70–89% inference, <50% speculation), and correct its own errors.
3. Self-Refine pipeline — generates corrected examples from the test bank and merges all sources into training_data_master.jsonl.
4. 10K eval bank — 8,999 questions across recon / exploit / defense / attribution / incident / compliance / psychology / tutor / whiterabbit / software-engineering.
5. Terminal UI — Parrot-OS two-line prompt, neon-northern-lights palette, ASCII "JEDI" banner, dynamic status bar, tool panel, White Rabbit Mode animation.

⏳ Left To Do (in order)

1. Run tune_jedi.py to produce jedi_lora_adapter/ (QLoRA, r=16, ~1 epoch over 20K examples).
2. Run eval_benchmarks.py on the fine-tuned model to get post-training numbers.
3. Convert the LoRA adapter to GGUF (so it can be applied to the GGUF base at inference) — see below.
4. Optionally re-run all the big public benchmarks (MMLU / MMLU-Pro / GPQA / IFEval / BFCL) and record the scores.
5. Apply adapter in the terminal — load model/ + LoRA at runtime.

---

How To Finish It (on your laptop)

1. Clone & install

git lfs install
git clone https://huggingface.co/FerrellSyntheticIntelligence/JEDI
cd JEDI
pip install -r requirements.txt
# If you want LoRA training:
pip install "transformers>=5.2.0" peft bitsandbytes datasets accelerate torch

2. Run the fine-tune (CPU or GPU)

# Quick smoke test first (100 examples, 20 steps):
python3 tune_jedi.py --quick

# Full run:
python3 tune_jedi.py

This produces jedi_lora_adapter/ (PEFT LoRA weights). On a laptop GPU this takes minutes; on CPU it can take hours. The script uses 4-bit QLoRA so VRAM/RAM stays low.

3. Apply the adapter to the GGUF for inference

Two options:

• Transformers path: python3 tune_jedi.py --apply loads base + LoRA and tests a generation.
• GGUF path (recommended for the terminal): convert the LoRA to a GGUF adapter and load it with llama_cpp:

  from llama_cpp import Llama
  llm = Llama(model_path="model/LFM2.5-1.2B-Instruct-Q4_K_M.gguf",
              lora_path="jedi_lora_adapter/adapter.bin")  # after conversion
  

4. Re-benchmark

python3 eval_benchmarks.py

Then compare to the pre-fine-tune baseline in eval_results.json.

---

Data Format

All training files are ShareGPT-style JSONL:

{"messages":[{"role":"system","content":"..."},{"role":"user","content":"..."},{"role":"assistant","content":"..."}],"domain":"machiavelli_psych"}

The tune_jedi.py script converts this to the LFM2.5 ChatML template (<|im_start|>...<|im_end|>) and masks the loss on non-assistant turns.

Domain distribution in training_data_master.jsonl

cybersec_exploit (1785), self_refine_* (≈7000 across coding/recon/defense/incident/attribution/psych/tutor/whiterabbit), veritas_* (723), machiavelli_psych (6769), cybersec_defense (1154), cybersec_recon (254), whiterabbit (343), cybersec_swarm (247), etc.

---

Architecture

User Input
  → Quadruflow Router (LOGICAL / FACTUAL / CREATIVE / PROCEDURAL)
  → Chain Amplifier (reasoning scaffold)
  → LFM2.5 1.2B Inference (llama-cpp or transformers)
  → VERITAS Loop (self-verify, confidence, correct)
  → Attestation Loop (3-check quality gate)
  → Memory Store (FAISS + Ebbinghaus decay)
  → JEDI Modules (swarm, legal gate, comms, ledger)
  → Response

The Veritas layer is what's new since the original JEDI release: the model now rates its own confidence and flags speculation instead of presenting guesses as fact.

---

Intended Use & Ethics

• Built for authorized defense only — all operations require legal authorization.
• Immutable audit ledger for every action.
• Built-in safety gate refuses harmful requests.
• Human-in-the-loop required for kinetic actions.

Citation

@misc{jedi2026,
  title={JEDI: Joint Entity Defense Infrastructure},
  author={Ferrell Synthetic Intelligence},
  year={2026},
  url={https://huggingface.co/FerrellSyntheticIntelligence/JEDI},
  license={MIT}
}

Contact

Ferrell Synthetic Intelligence — Neuro_Nomad