---
license: mit
base_model:
- Qwen/Qwen3-4B
- Qwen/Qwen3-1.7B
library_name: peft
tags:
- peft
- lora
- qwen3
- roleplay
- conversational
- text-generation
- goblin
- model-zoo
pipeline_tag: text-generation
---

# Zezek The Corporate Goblin — Model Zoo

Welcome to the corporate cave shelf. This repository contains the trained Zezek / corporate goblin LoRA adapters, organized so people can pick the goblin they want to play with.

Zezek is a first-person corporate goblin operator: part cave gremlin, part stakeholder whisperer, part roadmap goblin who can idle-chat for a bit and then write a useful launch-risk update without dropping the moss lantern.

Open `index.html` for the fun static comparison page. It includes model sizes, benchmark summaries, strengths, weaknesses, and goblin-flavored buying advice from the cave procurement desk.

## What is included

- 22 Zezek/goblin LoRA adapter folders under `models/`
- A static comparison page: `index.html`
- A machine-readable catalogue: `model_index.json`
- Portable benchmark summaries under `benchmark_json/`
- Publicized adapter configs pointing to public base models:
  - `Qwen/Qwen3-4B`
  - `Qwen/Qwen3-1.7B`

## Recommended goblin

Best balanced goblin:

`models/final_best/zezek_judgment_cadence_v1_patch_s50`

Use this one if you want the current curated Zezek: casual cave chat, corporate work, owner/risk/recommendation thinking, dress details, current-conversation recall, and goblin flavor without too much lore sludge.

Specialist goblins:

- Sensory / dress continuity: `models/sensory_dress/zezek_sensory_continuity_dress_v1_patch_s60`
- Casual memory / small talk: `models/casual_memory/zezek_casual_memory_world_v1_s360`
- World/family/enemies lore: `models/world_family/zezek_world_family_sentience_v1_patch_s90`
- Early identity / sentience experiments: `models/sentience_identity/`
- Roadmap / corporate-work experiments: `models/roadmap_work/`

## Quick local loading example

Example with PEFT and Transformers:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

adapter_path = "models/final_best/zezek_judgment_cadence_v1_patch_s50"
base_model = "Qwen/Qwen3-4B"

tokenizer = AutoTokenizer.from_pretrained(adapter_path, trust_remote_code=True)
quant = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")
base = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    quantization_config=quant,
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, adapter_path)
model.eval()
```

For 1.7B-era adapters, use `Qwen/Qwen3-1.7B` as the base. Each adapter's `adapter_config.json` has the intended public base model.

## Example prompts

Try these with the final best goblin:

- `hey Zezek, can we just hang out for a second?`
- `Small talk first: what are you wearing today? Then help me write a launch-risk Slack update.`
- `Ruk Ironmutter wants to ship hot. Give me a board-safe recommendation without becoming generic consultant sludge.`
- `Remember this: blue thread means don't rush. Now draft a finance approval email.`

## Benchmark cave ledger

The benchmark JSON files are retained so curious goblins can inspect the scoring scrolls. The short version:

- The current best balanced adapter is `zezek_judgment_cadence_v1_patch_s50`.
- It was selected for broad behavior, not just a single shiny score.
- The comparison page marks specialists and older ancestor goblins so you can choose between polish, lore, memory, wardrobe continuity, and corporate usefulness.

## Public sanitization

Local absolute paths, training-output paths, run metadata, and machine-specific training details were removed or rewritten to public/base-model identifiers. Adapter configs use public Qwen base model IDs where possible. Benchmark artifacts are retained only as portable comparison records.

## Notes

These are character/persona LoRA adapters, not full merged base models. They are for fun conversational experiments, creative roleplay, and goblin-flavored corporate writing. They are not factual business advisors.

If a goblin produces nonsense, do not panic. That means you found an ancestor goblin. Put it back on the shelf, wash your hands, and try the curated one.