jokegen2-1t-sft

LoRA adapter for Kimi-K2-Thinking. Fine-tuned on ~46k curated comedy examples from recent events (2026).

what it is

Supervised fine-tuning on scraped comedy data: tpot, standup transcripts, reddit. Teaches the model structure, voice, and the "turn" that tries to dig some relevant/recent in 2026 insight into the topic. It also teaches the model to use format tags to control the output style.

quickstart

pip install tinker transformers
export TINKER_API_KEY=your_key

import tinker
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("moonshotai/Kimi-K2-Thinking", trust_remote_code=True)
sampler = tinker.ServiceClient().create_sampling_client(
    model_path="tinker://c1788855-13a0-5d85-a45b-873b89e2e29c:train:0/sampler_weights/final"
)

prompt = "<|im_start|>system\nYou write sharp, witty comedy.<|im_end|>\n<|im_start|>user\nwrite a joke about startups<|im_end|>\n<|im_start|>assistant\n"
response = sampler.sample(
    prompt=tinker.types.ModelInput.from_ints(tokenizer.encode(prompt)),
    sampling_params=tinker.types.SamplingParams(max_tokens=256, temperature=0.8, stop=["<|im_end|>"]),
).result()
print(tokenizer.decode(response.sequences[0].tokens[len(tokenizer.encode(prompt)):]))

local inference

tbd