tinker-rl-arch_gsm8k_kimi-k2-kimi-k2

LoRA adapters trained with GRPO on top of moonshotai/Kimi-K2-Thinking using the Tinker cloud training service. Part of the TinkerRL-Bench release for our NeurIPS submission "A Unified Benchmark for RL Post-Training of Language Models" (repo).

Training configuration

Base model	moonshotai/Kimi-K2-Thinking
Experiment tag	arch_gsm8k_kimi-k2
Campaign	None
Task	gsm8k
Seed	42
LoRA rank	16
Learning rate	1e-05
Group size	4
Training steps	20
Platform	Tinker (tinker)
Training run ID	51a8ef9e-15ef-5f8f-bda1-78ee51387a12

Metrics

Metric	Value
Last-10 reward avg	0.8
Peak reward	1.0

Checkpoints in this repo

Step	Original Tinker URI	Local path
sampler_weights/final	tinker://51a8ef9e-15ef-5f8f-bda1-78ee51387a12:train:0/sampler_weights/final	final

How to load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "moonshotai/Kimi-K2-Thinking"
adapter = "arvindcr4/tinker-rl-arch_gsm8k_kimi-k2-kimi-k2"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, adapter, subfolder="final")  # or "<step>"

Companion releases

• Dataset: arvindcr4/tinker-rl-bench-wandb — all 334 W&B runs + 9,255 history rows
• Manifest: arvindcr4/tinker-rl-bench-checkpoints — full catalogue of every Tinker URI
• Code: pes-llm-research/tinker-rl-lab

Citation

@misc{tinkerrlbench2026,
  title   = {A Unified Benchmark for RL Post-Training of Language Models},
  author  = {Arvind, C. R. and Jeyaraj, Sandhya},
  year    = {2026},
  note    = {NeurIPS submission, https://github.com/pes-llm-research/tinker-rl-lab}
}

License

Apache 2.0. The underlying base model retains its original license — please check moonshotai/Kimi-K2-Thinking for any usage restrictions.