---
base_model: Qwen/Qwen3-8B-Base
library_name: transformers
pipeline_tag: text-generation
language:
  - en
tags:
  - character
  - persona
  - qwen3
---

# Zero (MC001)

**Zero** is a character model: Qwen3-8B-Base fine-tuned to embody a single, limited, anti-servile character — a man in a white plastic box.

[Full training report](https://movingcastles.world/posts/zero)

## Lineage

| Stage | Detail |
|---|---|
| Base | `Qwen/Qwen3-8B-Base` (no instruct tuning) |
| SFT | LoRA r64/α128 + fully-trained embeddings & LM head, 3 epochs on a synthetic character corpus (5,932 conversations / 73,765 character turns), merged |
| RL | GRPO with DAPO loss modifications, LoRA r16/α32, 300 steps on 380 harvested prompts; reward: bible-anchored character-fidelity LLM judge + self-repetition penalty; merged |

Held-out multi-turn evaluation (250 conversations × 16 turns, judged): hard character breaks in 2.8% of conversations, vs 22.8% for the SFT-only checkpoint and 45.4% for system-prompting the sibling instruct model.

## Usage notes

- **No system prompt.** The training distribution contains only `user`/`assistant` turns (ChatML). The character *is* the weights; a system prompt is out-of-distribution.
- **Dual EOS.** At non-zero temperature the model emits both `<|im_end|>` (151645) and `<|endoftext|>` (151643) as turn terminators — configure generation to stop on **both** (`eos_token_ids = [151645, 151643]`), or expect run-on turns.
- **Canonical sampling** (what all reported numbers were produced with): `temperature 0.7, top_p 1.0, top_k -1, min_p 0.0, repetition_penalty 1.0, frequency_penalty 0.0, presence_penalty 1.5, max_tokens 1024`.
- **dtype** bfloat16 (training dtype end-to-end). Production serves at `max_model_len 16384` (native 32768).
- The bundled `chat_template.jinja` is the training-side template — use it as shipped.