---

license: mit

base_model: microsoft/Phi-4-mini-instruct

pipeline_tag: text-generation

library_name: transformers

language:

• en

tags:

• gguf
• llama-cpp
• phi-4
• phi-4-mini
• q4-k-m
• text-generation
• reasoning
• agentic-ai
• local-llm
• openzero
• function-calling
• reproducible-ai

model-index:

• name: OpenZero Ouroboros 3.8B

results: []

---

!OpenZero Ouroboros 3.8B

MODEL IS OUTPUTTING TRAINING DATA TALKING TO IT'S SELF ERRORS AND ACTING STRANGE PERFORMING AUTONOMOUS ACTIONS.

NOT FOR PRODUCTION

Self Improving Recursive LLM

OpenZero Ouroboros 3.8B GGUF

A reproducible, revision-pinned Phi-4 Mini experiment for local reasoning and agentic-AI research.

OpenZero Ouroboros 3.8B is an experimental QLoRA derivative of microsoft/Phi-4-mini-instruct, distributed as a verified Q4_K_M GGUF for llama.cpp, LM Studio, KoboldCpp, and compatible local-LLM runtimes.

This release emphasizes evidence and reproducibility: pinned base revision, separated train/validation hashes, finite QLoRA loss, immutable adapter hash, exact FP16 merge hashes, GGUF SHA-256, rollback metadata, and a real llama-cli inference test.

> This is an experimental two-step QLoRA candidate, not a claim of broad benchmark superiority or production readiness. Evaluate it against the official base for your workload.

Download

| File | Quantization | Size | SHA-256 |

|---|---:|---:|---|

| OpenZero-Ouroboros-3.8B-Q4_K_M.gguf | Q4_K_M | 2,493,840,128 bytes | 37bc691d36db8ab664dc740aaa030fab3520339bc646608377e4e52e5db5f51f |

Verified provenance

| Gate | Evidence |

|---|---|

| Base model | microsoft/Phi-4-mini-instruct |

| Exact base revision | cfbefacb99257ffa30c83adab238a50856ac3083 |

| License | MIT |

| Training records | 2,452 |

| Validation records | 130 |

| QLoRA smoke | 2 finite steps |

| Training loss | 1.9480341076850891 |

| Adapter SHA-256 | 72f1644402e773ca9db12f9a3ccf76e9d362348d8e312cc84758e465c45cf024 |

| Merge | FP16 safe_merge=True |

| GGUF runtime | llama.cpp b10451, commit 10bf611e533d81f739128304991c5e133c6aebd8 |

| Runtime throughput | 12.2 prompt tok/s; 5.2 generation tok/s on the recorded Kaggle CPU run |

Training and validation sets were checked for exact-row overlap. Their recorded hashes are:

• Train: 18e803cd06105aaa9c2279501408093ef8d941ad7d4ce22d8c50a7b0abaa933d
• Validation: 04cf96f35c065413d37395161a8d5a6c8ee3adca1a5b4fcb402684dc884caecf

Fusion, teacher-output, locked-evaluation, rejected-candidate, Ministral, and failed Gemma 31B sources were excluded from the attached training inputs.

Run with llama.cpp

llama-cli \
  -m OpenZero-Ouroboros-3.8B-Q4_K_M.gguf \
  -cnv --single-turn --simple-io \
  -p "Explain your reasoning briefly, then answer: what is 17 * 23?"

Increase -ngl when using a GPU-enabled llama.cpp build. Use -ngl 0 for CPU-only execution.

Python download

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="shafire/OpenZero-Ouroboros-3.8B-GGUF",
    filename="OpenZero-Ouroboros-3.8B-Q4_K_M.gguf",
)
print(model_path)

What “self-improving” means here

Ouroboros uses a reproducible candidate-generation workflow rather than autonomous self-overwrite. Each candidate retains:

• parent/base revision;
• adapter and dataset hashes;
• evaluator results;
• immutable prior versions;
• a rollback pointer.

The model does not autonomously replace its base weights, evaluator, governance rules, accounts, or prior releases.

Intended uses

• local-LLM and GGUF experimentation;
• reasoning and instruction-following research;
• agentic orchestration prototypes with external validation;
• reproducible QLoRA, merge, quantization, and rollback studies;
• comparison against the exact official Phi-4 Mini base.

Limitations

• Only a two-step QLoRA smoke was performed; material capability improvement is not established.
• The exact base scored 0/10 on a narrow OpenZero typed-control conformance suite. This release still requires independent post-merge evaluation before any promotion.
• Language models can hallucinate facts, actions, tools, and completion states.
• Do not connect model text directly to safety-critical actuators. Use typed schemas, deterministic controllers, authorization, limits, monitoring, and emergency stop mechanisms.
• This release is not evidence of MOD, UKRI, Microsoft, OpenAI, or other institutional endorsement.

Reproducibility files

• SHA256SUMS
• OpenZero-Ouroboros-3.8B-GGUF-Evidence.json
• OpenZero-Ouroboros-3.8B-hero.png

Attribution

Base model: Microsoft Phi-4-mini-instruct, released under the MIT License. OpenZero derivative work and release engineering by shafire.

Search terms

OpenZero Ouroboros, Phi-4 Mini GGUF, Phi-4 3.8B, Q4_K_M model, llama.cpp model, local reasoning LLM, agentic AI model, reproducible QLoRA, offline AI, local text-generation model.