RWKV-Red-Team/ARWKV-R1-7B

ARWKV🪿

Paper Link👁️ | Github✅

ARWKV-R1-7B (Preview 0.1)

Preview version with RWKV-7 time mixing and Transformer MLP

📌 Overview

ALL YOU NEED IS RWKV

This is an early preview of our 7B parameter pure RNN-based model, trained on 2k context length (only stage-2 applied, without SFT or DPO) through 3-stage knowledge distillation from DeepSeek-R1-Distill-Qwen-1.5B. While being a foundational version, it demonstrates:

• ✅ RWKV-7's efficient recurrence mechanism
• ✅ No self-attention, fully O(n)
• ✅ Constant VRAM usage
• ✅ Single-GPU trainability

Roadmap Notice: We will soon open-source different enhanced versions with:

• 🚀 16k+ context capability
• 🧮 Math-specific improvements
• 📚 RL enhanced reasoning model

How to use

pip3 install --upgrade rwkv-fla transformers

from transformers import AutoModelForCausalLM, AutoTokenizer


model = AutoModelForCausalLM.from_pretrained(
    "RWKV-Red-Team/ARWKV-R1-7B",
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "RWKV-Red-Team/ARWKV-R1-7B"
)

system_prompt = "You are a world class trivia AI - provide accurate, succinct responses. "
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": prompt}]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
text = text + "<think>"
print(text)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

streamer = TextIteratorStreamer(tokenizer, skip_prompt=False, skip_special_tokens=False)


generation_kwargs = dict(model_inputs, streamer=streamer, max_new_tokens=8192, do_sample=True,tokenizer=tokenizer,stop_strings=["<｜end▁of▁sentence｜>"])
thread = threading.Thread(target=model.generate, kwargs=generation_kwargs)
thread.start()

print("Streaming output:")
for new_text in streamer:
    print(new_text, end="", flush=True)

thread.join()

The output looks like :

<｜begin▁of▁sentence｜>You are a world class trivia AI - provide accurate, succinct responses. <｜User｜>The world's largest rainforest, home to approximately three million species of plants and animals, is named after which river?<｜Assistant｜><think>
Okay, so I'm trying to solve this question about the world's largest rainforest and which river it's named after. Hmm, first, I think rainforest names often have links related to the region it's in. The most famous rainforest in the world is the Amazon. I remember hearing a lot about it being called that because rainforests are connected to specific river systems. 

Now, I'm trying to recall which river is named after the Amazon. I think it's the Amazon River. But I want to be sure. Let me see... the Amazon is a major rainforest located in South America. The Amazon River flows through it, which is why it's named after it. That makes sense because it's a very important river. I recall reading somewhere that all the rainforests are named after rivers related to their regions. So if the Amazon is named after its River, then the name would naturally be related to its source.

I wonder if it's the Amazon itself that's named after it, or another river named after it. But the official name for the Amazon is the Amazon Rainforest. The most significant rainforest in the world is the Amazon, and its name probably started with river-sounding names.
</think>

The largest rainforest located in South America is the Amazon. It is named after the river named after it, which is the Amazon River. Therefore, the Amazon River is the name given to the Amazon Rain Forest.

✈️ Benchmark

	Qwen2.5-7B-Instruct	ARWKV_7B	ARWKV_R1_7B
MMLU	71.72	62.41	67.25 ↗️
GSM8K	82.34	68.67	56.06 ↘️
WinoGrande	71.35	39.95	51.93 ↗️
IfEval	73.62	52.16	60.31 ↗️
Arc-c	54.86	52.22	29.01 ↘️

🔑 Key Features

Component	Specification	Note
Architecture	RWKV-7 TimeMix + SwiGLU	Hybrid design
Context Window	2048 training CTX	Preview limitation
Training Tokens	40M	Distillation-focused
Precision	FP16 inference recommended(16G Vram required)	15%↑ vs BF16

🏗️ Architecture Highlights

Core Modification Flow

Transformer Decoder Layer:
- Multi-head Latent Attention(MLA)
+ RWKV-7 Time Mixing (Eq.3)
- RoPE Positional Encoding
+ State Recurrence
= Hybrid Layer Output