---
license: apache-2.0
base_model: moonshotai/Kimi-K2-Thinking
tags:
  - comedy
  - lora
  - rl
  - grpo
---

# jokegen2-1t-rl

LoRA adapter for [Kimi-K2-Thinking](https://huggingface.co/moonshotai/Kimi-K2-Thinking). RL-tuned with comedy rubrics.

## what it is

Reinforcement learning on top of SFT. It uses a qualitative rubric similar to how Kimi was able to score high on eqbench. The rubrics score for specificity, turn, compression, and commitment. There is also one prescriptive rubric to prevent reward hacking, no laugh signaling, no explanation, no hedging, and no AI-isms -- otherwise the model will learn to append something similar to "lmao 💀" to everything.

For this experiment, I used qwen3 8b to grade the jokes. 

## quickstart

```bash
pip install tinker transformers
export TINKER_API_KEY=your_key
```

```python
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://4977bc84-470e-50e8-89fd-0bf9cd13c372:train:0/sampler_weights/000050"
)

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