PunTune 0.6 is a rank-32 LoRA adapter on Thinking Machines' Inkling (975B MoE, 41B active), trained with GRPO on Tinker to answer any topic with exactly one original dad joke. It's the checkpoint from the seventh and final reward design (the pre-registered "reader's-objection" rubric) in the DadBench project.

What training actually did

The jokes aren't the point. Measured on a frozen benchmark across the full thinking-effort dial, training didn't teach the model new wit — it taught it to pick one answer and stop. Base Inkling brainstorms brilliantly and fails at the commit (rambling into truncation); PunTune reaches the same punchlines at ~a quarter of the tokens.

Frozen-ladder score: 0.456 (base Inkling: 0.374 — an effort-matched lift of +0.082), on ~a quarter of base's tokens.
What RL bought: reliability and economy, not humor creation.
Training: GRPO (policy-gradient) via tinker-cookbook, rank-32 LoRA, thinking-effort 0.9, ~$200 total across seven reward designs.

Sibling checkpoint: aimatey/PunTune-0.5 (reward round R5) — statistically tied under both instruments; different reward, same model quality.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "thinkingmachines/Inkling"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, "aimatey/PunTune-0.6")

msgs = [
    {"role": "system", "content": "You are a master of dad jokes. When given a topic, reply with exactly one original dad joke about that topic. No preamble, no explanation."},
    {"role": "user", "content": "Topic: garden hoses"},
]
# ...generate as usual

Or serve the adapter with vLLM / SGLang via --lora-modules.

License

Apache-2.0, inherited from the Inkling base model. Please credit both PunTune and Thinking Machines' Inkling.

aimatey/PunTune-0.6

PunTune 0.6

What training actually did

Usage

Links

License

aimatey/
PunTune-0.6