bestie-genz-qwen2.5-1.5b-lora-gguf/README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
tags:
  - gen-z
  - persona
  - lora
  - gguf
  - llama.cpp
language:
  - en
pipeline_tag: text-generation
---

# bestie-genz-qwen2.5-1.5b-lora-gguf 🫶 — a Gen Z persona LLM

**bestie-genz-qwen2.5-1.5b-lora-gguf** is a small, local chat model that answers normal questions *correctly*
but talks like your chronically-online Gen Z best friend. It's
[Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) with a
lightweight LoRA adapter that reshapes **tone**, not facts.

Built to run fully offline on a MacBook Air M2 (16GB) via
[llama.cpp](https://github.com/ggml-org/llama.cpp) with Metal acceleration.

> ⚠️ **Slang snapshot:** This model reflects Gen Z slang and tone **as of 2026**.
> Internet language moves fast — some phrasing will feel dated over time. This is
> a point-in-time snapshot, not a living style guide.

---

## What's in this repo

| File | What it is | Size |
|------|-----------|------|
| `bestie-merged-q4_k_m.gguf` | Single-file merged model (base + bestie persona), easiest for LM Studio | ~940 MB |
| `qwen2.5-1.5b-instruct-q4_k_m.gguf` | Base model, quantized to Q4_K_M | ~986 MB |
| `bestie-lora.gguf` | The Gen Z LoRA adapter (kept **separate**, applied at runtime) | ~35 MB |
| `qwen2.5-1.5b-instruct-f16.gguf` | Full-precision base GGUF (optional) | ~2.9 GB |

The LoRA is intentionally **not merged** into the base weights. llama.cpp applies
it at inference time via `--lora`, so you can toggle the persona on/off with the
same base model.

---

## Base model & why

- **Base:** `Qwen2.5-1.5B-Instruct` (Apache-2.0)
- **Why:** Best-in-class reasoning/factuality for a sub-2B model, permissive
  license for redistribution, first-class llama.cpp conversion support, and small
  enough to LoRA-fine-tune and run on 16GB unified memory.

---

## Quick start (llama.cpp)

**Single merged model (no adapter needed):**

```bash
llama-cli \
  -m bestie-merged-q4_k_m.gguf \
  -st -n 200 --temp 0.7 \
  -p "Why is the sky blue?"
```

**CLI (single response):**

```bash
llama-cli \
  -m qwen2.5-1.5b-instruct-q4_k_m.gguf \
  --lora bestie-lora.gguf \
  -st -n 200 --temp 0.7 \
  -p "Why is the sky blue?"
```

**Server (OpenAI-compatible API):**

```bash
llama-server \
  -m qwen2.5-1.5b-instruct-q4_k_m.gguf \
  --lora bestie-lora.gguf \
  --host 127.0.0.1 --port 8080 -c 2048 --alias bestie
```

```bash
curl -s http://127.0.0.1:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"bestie","messages":[{"role":"user","content":"How do I stay motivated to work out?"}]}'
```

---

## Example outputs (base vs. bestie)

**Prompt:** *Why is the sky blue?*

- **Base:** "The sky appears blue because of a phenomenon called Rayleigh
  scattering. This happens when light from the sun travels through the Earth's
  atmosphere and is scattered..."
- **bestie:** "Bet, quick breakdown: Sunlight splits into colors, blue gets
  scattered more. No fluff, just signal."

**Prompt:** *Explain what a black hole is.*

- **Base:** "A black hole is a region of spacetime where gravity is so strong that
  nothing, not even light, can escape from it..."
- **bestie:** "Real talk, no BS: A black hole is a region of spacetime with such
  strong gravitational pull that nothing, not even light, can escape from it."

**Prompt:** *How do I stay motivated to work out?*

- **bestie:** "Real talk, no BS: Break your workout into manageable 30-minute
  sessions, track your progress, and always focus on form over intensity. Progress
  over perfection, always."

---

## Intended use

- Casual Q&A, study help, explanations, and everyday assistance with a fun,
  informal Gen Z voice.
- Great for demos, personal projects, and learning the LoRA → GGUF → llama.cpp
  pipeline.

### Out of scope / limitations

- **Small model (1.5B):** can hallucinate or oversimplify. Verify anything
  important.
- **Persona artifacts:** trained on a small (~165-example) dataset, so it
  occasionally over-uses signature phrases ("real talk, no BS", "drop your
  context") or leaks a stray token. Larger/cleaner data would smooth this out.
- **English only**, informal register — not suitable for formal/professional
  writing.
- **Not safety-tuned** beyond the base model's alignment.

---

## Training details

- **Method:** LoRA (PEFT) — the *only* step done in `transformers`/`peft`, since
  llama.cpp is inference-only and can't train.
- **Everything else via llama.cpp:** GGUF conversion (`convert_hf_to_gguf.py`),
  LoRA→GGUF export (`convert_lora_to_gguf.py`), quantization (`llama-quantize`),
  inference (`llama-cli`), and serving (`llama-server`).
- **Data:** ~165 hand-written (prompt, response) pairs — normal questions answered
  correctly in an authentic Gen Z voice — plus ~20 validation examples.
- **Hyperparameters:** low rank (small-dataset appropriate), few epochs, tuned for
  16GB unified memory on Apple Silicon (MPS).

---

## License

Apache-2.0, inherited from the Qwen2.5-1.5B-Instruct base model.