---
license: apache-2.0
language:
- en
tags:
- recursive-language-model
- hybrid-mind
- causal-lm
- multimodal
- self-automated
- reinforcement-learning
- continual-learning
- memory-augmented
pipeline_tag: text-generation
library_name: transformers
model_type: sentience_cascade
---

# Sentience.Cascade.II

**Recursive Language Model (RLM) · Hybrid Mind Frame**
**1.147B Parameters · 64K Context Window · Dual T4 Trained**

---

## Overview

**Sentience.Cascade.II** is not a Large Language Model (LLM).  
It is a **Recursive Language Model (RLM)** — a novel architecture where every
forward pass includes multiple self-recursive refinement steps, episodic short
and long-term memory, and a fully wired Hybrid Mind module that runs *as one
integrated frame*, not as sequential pipeline stages.

All cognitive subsystems operate inside a single unified forward pass.

---

## Architecture

| Component | Detail |
|---|---|
| Architecture type | Recursive Language Model (RLM) |
| Parameters | ~1.147B |
| Context window | 64,000 tokens |
| Attention | Grouped Query Attention (16 heads / 4 KV heads) |
| Positional encoding | RoPE (θ=500,000) |
| FFN | SwiGLU |
| Normalisation | RMSNorm |
| Weight format | safetensors (float32 on disk, bfloat16 for training) |
| Vocabulary | 65,536 (BPE ByteLevel) |

---

## Hybrid Mind Frame — Self-Automated (S.A.) Modules

All modules are active simultaneously inside each transformer layer.
None are optional pipeline steps — they are weights baked into the model.

| Module | Role |
|---|---|
| S.A. Meta Learning Gate | Scales activation magnitude as a proxy learning signal |
| S.A. Reinforcement Learning Head | Scalar reward prediction per forward pass |
| S.A. Continual Learning Gate | Soft forgetting-protection via decay gates |
| S.A. Adaptive Learning Scale | Per-token hidden-state scaling |
| S.A. Rewrite Gate | Token-level hidden-state rewriting delta |
| S.A. NLP Head | Span boundary logits for structured extraction |
| S.A. Problem Solving Head | 8-class step-type classification |
| S.A. Innovation Noise | Trainable exploration noise (active during training only) |
| S.A. Debug Probe | 4-class anomalous activation detector |
| S.A. Advanced Short-Term Memory | 512-slot episodic rolling buffer |
| S.A. Advanced Long-Term Memory | 1024-slot consolidated episodic store |
| S.A. Recursive Seed Learning | Multi-step (×4) recursive refinement loop |
| S.A. Self-Evaluation & Reward | Scalar self-score head |
| S.A. Goal & Constraint Engine | Residual goal-projection delta |
| S.A. Memory Consolidation | Automatic STM→LTM every 8 layers |
| S.A. Introspection Interface | 64-dim interpretable summary of hidden state |
| S.A. Recursive Outer Loop Gate | Final gate before residual output |
| Conversational Intelligence | 32-class dialog-act classification head |
| MultiModal (Text/Image/Audio/Video) | Linear projection from ViT-L / mel-spec / video dims |

---

## Recursive Language Model Core

Unlike a standard transformer that processes tokens once per layer, **Sentience.Cascade.II**
applies a **RecursiveSeedLayer** after all transformer blocks.  This layer runs
`num_recursive_steps=4` passes of attention + FFN with a shared-weight inner loop,
allowing the model to internally "think again" before producing logits.

This is the defining feature of the RLM architecture:
> *Output is not produced after one pass — it is refined recursively.*

---

## Memory System

- **Short-Term Memory (512 slots):** Updated every forward pass via a write gate.  
  Cross-attended by every layer, giving the model persistent intra-context state.
- **Long-Term Memory (1024 slots):** Consolidated from short-term every 8 layers via
  a separate consolidation gate with 0.99/0.01 EMA blend.  
  Persists across training steps when fine-tuning.

---

## Multimodal Support

Three input projection heads accept external embeddings:

| Modality | Input dim | Projection |
|---|---|---|
| Image | 1024 (ViT-L patch) | Linear → 2048 |
| Audio | 128 (mel-spectrogram) | Linear → 2048 |
| Video | 1024 (frame embedding) | Linear → 2048 |

These are additive prefix embeddings — concatenate modality tokens before input_ids.

---

## Chat Template

```
<|system|>You are Sentience.Cascade.II, a recursive reasoning model.
<|user|>What is consciousness?
<|assistant|>
```

---

## Fine-Tuning

This is the **base pretrained initialisation** — weights are randomly initialised
and the tokenizer is bootstrapped.  Fine-tune on your domain corpus using standard
causal-LM training.

Recommended fine-tune config:

```python
from transformers import TrainingArguments

args = TrainingArguments(
    output_dir           = "./sc2-finetuned",
    per_device_train_batch_size = 1,
    gradient_accumulation_steps = 16,
    num_train_epochs     = 3,
    learning_rate        = 2e-4,
    lr_scheduler_type    = "cosine",
    warmup_ratio         = 0.03,
    bf16                 = True,
    gradient_checkpointing = True,
    save_strategy        = "steps",
    save_steps           = 500,
    logging_steps        = 10,
    report_to            = "none",
)
```

> **Note:** Because `SentienceCascadeModel` is a custom architecture, you will
> need to register it with the HuggingFace `AutoModel` registry or load it
> with `trust_remote_code=True` after placing the model code in the repo.

---

## Citation

```bibtex
@misc{sentiencecascade2,
  author       = {GODsStrongestSoldier},
  title        = {Sentience.Cascade.II: A Recursive Language Model with Hybrid Mind Frame},
  year         = {2025},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/GODsStrongestSoldier/Sentience.Cascade.II}},
}
```

---

## License

Apache 2.0