A 56M‑parameter, from‑scratch recurrent language model that combines a per‑channel leaky integrator (FWKV) with the RWKV‑8 ROSA copy‑signal mechanism. It is a research experiment designed to explore how far purely recurrent architectures can go on small‑scale, curated conversational data.

The model was chat‑tuned on HuggingFaceH4/ultrachat_200k and uses a simple two‑role template:

<|user|> Your message
<|assistant|> Model reply<|endoftext|>

Model Description

Architecture: 14 stacked FWKV blocks. Each block replaces the standard attention with a fixed, data‑independent decayed accumulator: stateₜ = W·stateₜ₋₁ + kₜ·vₜ, where W = clamp(sigmoid(w), min=0.1). The recurrence is computed exactly via a vectorised parallel scan (no approximations).
ROSA (Rapid Online Suffix Automaton): A parameter‑free, causal predictor that injects the token that historically followed the longest matching suffix of the current context. The ROSA signal is embedded and added to the input representation of each token.
Factorised embedding/head: 128‑dimensional embedding space, projected to a 512‑dimensional model space, with tied weights.
Context length: 1024 tokens. No positional embeddings are used, making the context window “free” in terms of parameters.
Tokenizer: GPT‑2 tokenizer extended with the special tokens <|user|> and <|assistant|>.

Architecture graph for FWKV/FWKV-ROSA. Open in hfviewer

Uses

Direct Use

FWKV-ROSA is intended for research on efficient language models and for educational demonstrations of recurrent architectures. You can chat with it in a multi‑turn setting using the template above.

Out‑of‑Scope Use

This model is not suitable for any production or safety‑critical application.
It has not been aligned with RLHF or other safety methods and may generate inappropriate or harmful content.
The limited size and training data mean it cannot be relied upon for factual knowledge or reasoning.

Bias, Risks, and Limitations

Trained on a relatively small synthetic dataset, the model can produce repetitive or nonsensical output.
The ROSA copy mechanism may occasionally copy large chunks of the user’s prompt verbatim.
Biases present in the original UltraChat data are likely reflected in the model’s responses.

How to Get Started

The model relies on a custom architecture. To load it, you must provide the modeling_fwkv.py file (found in the repository) and trust the remote code:

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "FlameF0X/FWKV-ROSA",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("FWKV/FWKV-ROSA")

Then format your prompts exactly with the chat tokens:

python

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device).eval()

prompt = "<|user|> What is the capital of France?\n<|assistant|>"
input_ids = tokenizer.encode(prompt)
# ROSA IDs must be computed – you can import `rosa` from modeling_fwkv
from modeling_fwkv import rosa
rosa_ids = torch.tensor([rosa(input_ids)], device=device)
out = model(input_ids=torch.tensor([input_ids], device=device),
            rosa_ids=rosa_ids, use_cache=True)
# Continue autoregressive sampling…

For a fully working chat demo, see the Gradio app provided in the repository.

Training Details

Dataset

Name: UltraChat 200k
Splits: train_sft (20,000 examples), test_sft (1,000 examples)
Format: multi‑turn conversations; only assistant tokens contribute to the loss.

Training Procedure

Hyperparameter	Value
Architecture	14 FWKV blocks, d_model=512, d_emb=128
FFN multiplier	4
WKV decay floor	0.1
Batch size (per GPU)	8
Gradient accumulation	4
Effective batch size	32
Learning rate	3×10⁻⁴ (cosine schedule)
Weight decay	0.1
Gradient clipping	1.0
Optimizer	AdamW (fused)
Precision	bfloat16 mixed
Epochs	2
Hardware	1× NVIDIA T4 (15 GB)
Training time	~2 hours
Speed	~80,000 tokens/second

Gradient checkpointing was enabled to fit the 1024‑token sequences in memory.

Evaluation

Metric	Value
Validation loss	4.216
Validation perplexity	67.78

The perplexity is relatively high due to the small model size and limited training data. It is comparable to other similarly‑sized recurrent LMs on UltraChat.

Environmental Impact

The training ran for about 2 hours on a single NVIDIA T4 GPU (maximum power draw ~70 W), resulting in an estimated 0.14 kWh of electricity consumption and approximately 0.06 kg CO₂eq (assuming a grid carbon intensity of 0.4 kg/kWh). This is a negligible footprint.

Technical Specifications

Model type: Recurrent neural network (linear RNN)
Parameters: 56.2 million
- Backbone (FWKV blocks + embeddings): ~50M
- ROSA embedding: ~6.2M
Checkpoint format: PyTorch safetensors
Required files in the repo:
- config.json
- model.safetensors (or pytorch_model.bin)
- modeling_fwkv.py
- tokenizer.json / vocab.json / merges.txt
Auto‑mapping: The config.json includes "auto_map": { "AutoModelForCausalLM": "modeling_fwkv.FWKVLanguageModel" }, so loading with trust_remote_code=True will automatically locate the correct class.

Citation

If you use FWKV-ROSA in your research, please cite it as:

bibtex

@misc{fwkv-rosa,
  author = {FlameF0X},
  title = {FWKV-ROSA: A 56M Recurrent Chat LM with RWKV-style Decay and ROSA Copy Signal},
  year = {2025},
  howpublished = {\url{https://huggingface.co/FWKV/FWKV-ROSA}},
}

Additional Information

This model was built as an experiment to test the combination of a simple leaky integrator with the ROSA copy signal on a small, clean conversational dataset. It demonstrates that a pure linear RNN can learn to produce coherent multi‑turn dialogue without any attention mechanisms. Feedback and contributions are welcome!