Update README.md

README.md

@@ -17,9 +17,10 @@ tags:
 - quantization
 - ggml
 - llm
+- spooky-fun
+- haunted-ai
 ---
-
-# Halloween2B-Poltergeist v1.0 (GGUF)
+# Halloween2B-Poltergeist v1.0 (GGUF) – Now With Extra Ghosts!
 
 [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
 [![Format: GGUF](https://img.shields.io/badge/Format-GGUF-orange)](#quantized-files)
@@ -28,69 +29,61 @@ tags:
 [![Hugging Face](https://img.shields.io/badge/Hugging%20Face-Model-yellow)](https://huggingface.co/ghostai1/Halloween2B-Poltergeist)
 [![LinkedIn](https://img.shields.io/badge/LinkedIn-ccengineering-0A66C2?logo=linkedin&logoColor=white)](https://www.linkedin.com/in/ccengineering/)
 
-**Halloween2B-Poltergeist v1.0** is a compact seasonal model tuned for **spooky, cinematic, and playful** text generation.  
-This repository provides **GGUF** builds optimized for **local inference** with **llama.cpp** and compatible runtimes.
+**Halloween2B-Poltergeist v1.0** is your go-to compact model for all things **spooky, cinematic, and downright playful** text generation. Think of it as that mischievous ghost in your machine, rattling chains and whispering eerie tales just in time for the witching hour. We've got GGUF builds ready for **local inference** via **llama.cpp** – perfect for when you want to summon AI without summoning the cloud.
 
-**Quick links**
-- Model: https://huggingface.co/ghostai1/Halloween2B-Poltergeist
-- Runtime: https://github.com/ggerganov/llama.cpp
-- Author (LinkedIn): https://www.linkedin.com/in/ccengineering/
+**Quick Haunts (Links):**
+- Model Tomb: https://huggingface.co/ghostai1/Halloween2B-Poltergeist
+- Summoning Ritual: https://github.com/ggerganov/llama.cpp
+- Mad Scientist (Author): https://www.linkedin.com/in/ccengineering/
 
 ---
+## The Haunting Overview
+This bad boy is all about **tone and personality** – tuned to crank out Halloween vibes like a possessed typewriter. It's got that atmospheric drip, killer pacing, and voices that could make a skeleton shiver. Factual accuracy? Pfft, who needs it when you're brewing potions of pure imagination?
 
-## Overview
-
-This is a **tone/personality-focused** model geared toward Halloween-themed writing and roleplay. It aims for atmosphere, pacing, and character voice over strict factual accuracy.
+**Summon It For:**
+- Bone-chilling microfiction, creepy campfire yarns, and those urban legends that keep you up at night
+- Halloween roleplay (keeps it PG-13 unless you poke the bear... or ghost)
+- Narrating haunted mansions, spooky quests, and chit-chat with undead NPCs
+- Quick demos on your dusty old rig – low resources, high scares!
 
-Recommended for:
-- Horror microfiction, campfire stories, and urban legends
-- Halloween roleplay (PG-13 by default)
-- Haunted-house narration, quests, and NPC dialogue
-- Local demos and low-resource systems
-
-Not optimized for:
-- Complex reasoning
-- Factual retrieval and citations
-- Long-horizon planning and tool use
+**Leave It In The Crypt For:**
+- Brain-busting puzzles or deep math (it's more boo than IQ)
+- Fact-checking missions or pulling up Wikipedia (ghosts don't cite sources)
+- Epic sagas or AI that calls tools – this one's for short, sharp shocks
 
 ---
-
-## Model Details
-
-- **Model name:** Halloween2B-Poltergeist v1.0
-- **Base model:** Qwen2.5-2B-Instruct
-- **Fine-tuning:** LoRA (adapter-based; release may be merged depending on build)
-- **Language:** English
-- **Format:** GGUF (llama.cpp compatible)
-- **License:** MIT (base model license applies)
-
-### Training Snapshot (from run log)
-
-- **Trainer:** Unsloth SFTTrainer (new-style)
-- **Unsloth version:** 2026.1.4
-- **LoRA:** r=64, alpha=64
-- **Patched layers:** 28 (QKV=28, O=28, MLP=28)
-- **Examples:** 189
-- **Epochs:** 3
-- **Total steps:** 36
-- **Batch:** 2
-- **Grad accumulation:** 8
-- **Effective batch:** 16
-- **Precision:** bf16=True (fp16=False)
-- **Trainable params:** 73,859,072 / 1,617,573,376 (4.57%)
-- **Train loss:** 2.5160
-- **Runtime:** 31m 31s
-- **Output adapter:** `trainmodel/runs/halloween2b/adapter_latest`
+## Spectral Specs (Model Details)
+- **Model Moniker:** Halloween2B-Poltergeist v1.0
+- **Base Haunt:** Qwen2.5-2B-Instruct
+- **Tuning Tricks:** LoRA (adapter-style; might be merged in the final brew)
+- **Tongue:** English (with a dash of eldritch whispers)
+- **Format:** GGUF (plays nice with llama.cpp)
+- **Curse (License):** MIT (but the base model's rules still lurk)
+
+### Training Terror Log
+- **Trainer Ghoul:** Unsloth SFTTrainer (fresh off the slab)
+- **Unsloth Vintage:** 2026.1.4 (future-proofed for maximum spook)
+- **LoRA Lore:** r=64, alpha=64
+- **Patched Phantoms:** 28 layers (QKV=28, O=28, MLP=28 – fully possessed)
+- **Victim Count (Examples):** 189
+- **Epochs of Torment:** 3
+- **Steps to Madness:** 36
+- **Batch Brew:** 2
+- **Grad Graveyard:** 8 accumulations
+- **Effective Elixir:** Batch size 16
+- **Precision Potion:** bf16=True (fp16? Nah, we're floating point ghosts)
+- **Trainable Torments:** 73,859,072 / 1,617,573,376 (4.57% – lightweight horror)
+- **Final Scream (Train Loss):** 2.5160
+- **Ritual Runtime:** 31m 31s (quicker than a witch's broom)
+- **Spawn Point:** `trainmodel/runs/halloween2b/adapter_latest`
 
 ---
+## Quantized Crypts (Files)
+All these spectral files come from the same unholy source – just quantized for your mortal hardware. Pick your poison!
 
-## Quantized Files
+> Note: Swap in your real file names, you lazy necromancer.
 
-All files are produced from the same model and differ only in quantization.
-
-> Replace the filenames below with your actual release artifact names.
-
-| File | Quantization | Approx. Size |
+| File | Quantization | Approx. Size (in Soul GB) |
 |---|---|---:|
 | `release_v1.TQ1_0.gguf` | TQ1_0 | ~0.47 GB |
 | `release_v1.Q2_K.gguf` | Q2_K | ~0.68 GB |
@@ -103,23 +96,18 @@ All files are produced from the same model and differ only in quantization.
 | `release_v1.Q6_K.gguf` | Q6_K | ~1.27 GB |
 | `release_v1.Q8_0.gguf` | Q8_0 | ~1.65 GB |
 
-### Recommendations
-
-- **Default (balanced):** `Q4_K_M`
-- **Higher quality:** `Q5_K_M`, `Q6_K`, `Q8_0`
-- **Low RAM systems:** `Q3_K_M`, `Q2_K`
-- **Ultra-low memory (experimental):** `TQ1_0`
+### Summoning Suggestions
+- **Balanced Brew (Default):** `Q4_K_M` – Solid spooks without draining your ectoplasm.
+- **Premium Phantoms:** `Q5_K_M`, `Q6_K`, `Q8_0` – For when you want crystal-clear chills.
+- **Budget Banshees:** `Q3_K_M`, `Q2_K` – Runs on a potato, still scares the pants off.
+- **Micro-Malevolence:** `TQ1_0` – Experimental, for when your rig is basically a Ouija board.
 
 ---
-
-## Usage (llama.cpp)
-
-### CPU-only
-
+## Invocation Incantation (Usage with llama.cpp)
+### Pure Mortal (CPU-Only) Mode
 ```bash
 ./llama-cli \
   -m release_v1.Q4_K_M.gguf \
   -ngl 0 \
   -c 4096 \
-  -p "Tell a spooky but playful Halloween story set in a small town. Keep it PG-13, vivid, and cinematic."
-
+  -p "Tell a spooky but playful Halloween story set in a small town. Keep it PG-13, vivid, and cinematic."