ghostai1/halloween1.3b_poltergeist

Halloween2B-Poltergeist v1.0 (GGUF) – Now With Extra Ghosts!

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 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?

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!

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

---

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!

Note: Swap in your real file names, you lazy necromancer.

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
release_v1.Q3_K_S.gguf	Q3_K_S	~0.76 GB
release_v1.Q3_K_M.gguf	Q3_K_M	~0.82 GB
release_v1.Q4_K_S.gguf	Q4_K_S	~0.94 GB
release_v1.Q4_K_M.gguf	Q4_K_M	~0.99 GB
release_v1.Q5_K_S.gguf	Q5_K_S	~1.10 GB
release_v1.Q5_K_M.gguf	Q5_K_M	~1.13 GB
release_v1.Q6_K.gguf	Q6_K	~1.27 GB
release_v1.Q8_0.gguf	Q8_0	~1.65 GB

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.

---

Invocation Incantation (Usage with llama.cpp)

Pure Mortal (CPU-Only) Mode

./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."