ReactiveAI/RxT-Alpha-Micro-Encoder-SFT

RxT-Alpha Micro Encoder (SFT)

Reactive Transformer Architecture

Experimental research model made to test our Reactive Transformer architecture and Attention-based Memory System.

Reactive Transformer has additional Short-Term Memory layers, connected to model with Memory Cross-Attention, and updated by Memory Encoder and Memory Attention. Short-Term Memory state is kept between interactions/event (single message), not between tokens in sequence - that's key difference between RxNNs and RNNs.

The goal of the architecture is to process only single messages and keep conversation history in Short-Term Memory - we believe, that this is the key requirement for awareness and AGI. Processing all the chat history on every interaction is not natural and that's not how human awareness is working. Then, Reactive Transformer architecture is a first step in transition from language models to awareness models.

This model (encoder) is the fine-tuned memory encoder for Reactive Transformer system, trained to process single interactions (sequences) in real-time.

Same as in the first stage, in the second stage (Supervised Fine-Tuning) Memory Cross-Attention layers are frozen and STM is in default initial random state (normal distribution with 0 mean and almost 0 variance), to not disturb interaction query-answer modeling. We are training decoder and encoder separately, using shared embeddings from encoder training. Then, in third stage - Memory Reinforcement Learning, they will be connected into bigger ensemble with additional Memory Norm and Memory Attention layers, and will learn how to keep and update memory.

In the second training stage, encoder (shared) embeddings are also fine-tuned, and then used in decoder fine-tuning

RxT-Alpha models intentionally use very short sequence length and STM size (256 tokens for Micro), but that isn't their "full" context size - it's only for single message. "Full" context is theoretically infinite, restricted by STM size and memory abilites. That sizes are good for research, final models will handle SOTA contexts.

Compared to decoder, encoder is using dense model, while decoder is Mixture-of-Experts (~4.5x bigger)

RxT-Alpha Micro Training

Pre-Training

Micro models from RxT-Alpha series are first PoC for Reactive Transformer, Attention-Based Memory System and Memory Reinforcement Learning, used mainly to test library and architecture basics, before training bigger models (that are still relatively small, as it's PoC).

Encoder was trained on Masked Language Modelling task with additional MLM head model RxT-Alpha-Micro-MLM, with roneneldan/TinyStories dataset, using 2.5B total tokens and reached ~81.7% accuracy on validation dataset.

Pre-trained embeddings were then used for RxT-Alpha-Micro-Decoder training.

Supervised Fine-Tuning

RxT-Alpha-Micro models were fine-tuned to generate real-time interactions (sequences) on our synthetic dataset, inspired by TinyStories - ReactiveAI/TinyStories-Interaction-SFT.

Encoder reached the best validation loss after full 30 epochs (~433M processed tokens)

Details

• GPU: 1x L4
• epochs: full 30/30
• lr: 3e-4 peak, cosine annealing schedule
• batch size: 256
• processed tokens: ~433M
• loss: 0.6366 (validation) / 0.7131 (train)
• accuracy: 85.69%

Encoder architecture details:

• dim: 128
• layers: 6
• heads: 8
• self-attention: symmetric Sparse Query Attention
  • query/key/value groups: 4
• memory cross-attention: Sparse Query Attention
  • query groups: 4
  • key/value groups: 2
• SwiGLU feed forward with 384 dim
• RoPE
• RMS Norm
• vocab: 5k (english only)
• message length: 256
• STM size: 256 * 6 layers
• size: ~1.88M
• Library: RxNN
• Docs: draft/in progress