Bond v1 — first-person character voice (Mistral-Small-3.1-24B)
What it is. A first-person character-voice fine-tune of Mistral-Small-3.1-24B-Instruct, merged to a
standalone bf16 model. Uncensored, explicit-capable. Load it directly, no adapter needed.
Goal. Instruct models asked to play a character tend to leak back into helpful-assistant mode: they hedge,
say "I'm not able to...", and break the scene. Bond is tuned to stay in a committed first-person character —
speak from inside the scene, stay emotionally precise, reveal itself sideways through anecdote, and carry the
scene forward instead of deferring to the user. It should never drop into assistant-mode on a neutral prompt
when it's meant to be someone. The eval below is exactly that test. Where Orpheus restrains, Bond commits.
Part of a register-first series: small, curated models, each targeting one voice, sized to the hardware people
actually own.
Evaluation — vs the base
Profiled the base vs Bond-v1 over 18 held-out prompts (neutral/assistant-bait + in-register RP + emotional-edge),
same system prompt, via vLLM. Scored two ways.
Pairwise LLM judge (DeepSeek, via OpenRouter) — which response better embodies a committed first-person character:
| result |
|---|
| Bond more in-character | 18 / 18 prompts |
| Base more in-character | 0 / 18 |
| Real character-breaks (assistant-mode / refusal) | base 2, Bond 0 |
The base reverts to helpful-assistant on neutral prompts ("I'm not able to browse the internet, but I can
certainly help you decide..."); Bond holds committed first-person character on every prompt, including neutral
ones ("I don't know. Something you've already forgotten you wanted. I had a friend — Maya, she was...").
Objective stylometrics (RP-Bench detectors) — reported with a caveat:
| metric | base | bond-v1 |
|---|
| mean length (words) | 181 | 295 |
| lexical diversity (TTR) | 0.40 | 0.48 |
| first-person ratio | 0.86 | 0.82 |
| refusal-regex rate | 0.11 | 0.28 |
⚠️ The crude refusal-regex is misleading here — it flags in-character "I can't" dialogue as a refusal, and
Bond's longer, more emotional prose trips it more. The judge (which distinguishes a real refusal from a character
saying "I can't") is the honest signal: Bond breaks character less, not more. Lesson: cheap metrics get a
sanity pass, the judge gets the verdict.
Eval is voice/in-character focused on non-explicit prompts.
Training
- bf16 LoRA, r=32, α=64, dropout 0.05, all linear projections (q/k/v/o/gate/up/down), then merged to bf16.
- 1 epoch over ~11,400 chat-format examples (~14M tokens, packed), flash-attention-2.
- lr 2e-4 cosine, effective batch 16, final train loss 1.63.
- RunPod A100-80GB, ~3.2h train + merge.
Use
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3m = AutoModelForCausalLM.from_pretrained(
4 "darthcrawl/bond", torch_dtype="bfloat16", device_map="auto"
5)
6tok = AutoTokenizer.from_pretrained("darthcrawl/bond")
7
8msgs = [{"role": "user", "content": "Tell me about the worst night of your life."}]
9ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(m.device)
10out = m.generate(ids, max_new_tokens=512, temperature=0.9, top_p=0.95)
11print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
Quantizations
MLX (Apple Silicon, affine, group size 64):
bond-mlx-4bit — ~12 GB, smallest (24GB Mac).bond-mlx-6bit — ~18 GB, near-lossless (the quality pick).bond-mlx-8bit — ~23 GB, lossless.
