FlameF0X/i3-200m-v2

i3-200M - Hybrid Architecture Language Model

Model Description

The i3-200M Model (aka Redherring) is a novel hybrid architecture combining convolutional/recurrent layers with full attention layers for efficient language modeling. This architecture uniquely blends RWKV-style time-mixing with Mamba state-space dynamics in the early layers, followed by standard multi-head attention in deeper layers.

To use the model try it here.

Model Statistics

• Total Parameters: ~169.85M
• Architecture: 10 Hybrid (RWKV-Mamba) + 6 Full Attention Layers = 16 Total Layers
• Vocabulary Size: 32,000 tokens
• Hidden Dimension (d_model): 512
• Attention Heads: 16
• State Dimension (d_state): 32
• Max Sequence Length: 256
• Tokenization: BPE

Architecture Breakdown

Layers 1-10:  RWKV-Mamba Hybrid Blocks (Recurrent/Conv)
              ├─ RWKVMambaHybrid (Time-mixing + State-space)
              └─ Feed-Forward Network (4x expansion)

Layers 11-16: Full Attention Blocks
              ├─ Multi-Head Attention (16 heads)
              └─ Feed-Forward Network (4x expansion)

Comparison with i3-80M

Feature	i3-22M	i3-80M	i3-200m (This Model)
Parameters	22.6M	82.77M	169.85M
Architecture	24 Hybrid Layers	10 Hybrid + 6 Attention Layers	10 Hybrid + 6 Attention Layers
Hidden Dimension	512	512	512
Vocabulary Size	4,466	35,560	32,000
Training Dataset	TinyChat only	TinyStories + TinyChat + HQ Sentences	TinyStories + TinyChat + HQ Sentences + Wikitext
Total Tokens	~1M conversations	~3M+ tokens	N/A
Final Loss	~2.0	~2.0	1.6
Final Perplexity	7.29-9.70	7.29-10.0	5.2
Training Time	~17 hours	~2-4 hours	~1-2 hours
Attention Layers	None (Pure Hybrid)	6 Full Attention Layers	6 Full Attention Layers

Key Improvements Over i3-80M

To Be added

Key Features

1. Hybrid Architecture: Combines the efficiency of recurrent/convolutional processing with the power of attention

   • Early layers use RWKV-Mamba hybrid for efficient sequence processing
   • Later layers use full multi-head attention for complex pattern recognition

2. Memory-Optimized Training:

   • Streaming vocabulary building (no full text storage)
   • Vocabulary caching (build once, reuse)
   • Efficient chunk frequency counting
   • Automatic memory cleanup

3. Multi-Dataset Pre-training: Trained on diverse text sources for robust language understanding

   • TinyStories: Narrative and storytelling
   • TinyChat: Conversational dynamics
   • High-Quality English Sentences: Linguistic diversity
   • wikitext

4. BPE Tokenization:

   • Total tokens processed: N/A
   • Handles unknown tokens gracefully with , , , token

Training Details

Training Configuration

• Datasets:
  • agentlans/high-quality-english-sentences
  • roneneldan/TinyStories
  • starhopp3r/TinyChat
  • wikitext
• Training Steps: 250 iterations
• Batch Size: 4 (with gradient accumulation support)
• Learning Rate: 4e-4 (with warmup and cosine decay)
• Optimizer: AdamW with gradient clipping (max norm: 1.0)
• Hardware: NVIDIA P100 (16GB VRAM)
• Training Time: ~1-2 hours
• Framework: PyTorch

Training Dynamics

• GPU Utilization: Stable at ~N/A% during training
• GPU Memory: ~20% allocated (~4GB / 12GB) (Math is not mathing??)
• Power Usage: ~250W average
• Throughput: ~300 tokens/sec

Performance Metrics

Metric	Initial	Final
Training Loss	~10.0	1.6
Perplexity	~4000+	5.2

Technical Innovations

1. RWKV-Mamba Hybrid Recurrence: Combines RWKV's time-mixing with Mamba's state-space dynamics

   • Linear complexity for long sequences
   • Efficient recurrent processing
   • State-space modeling for temporal dependencies

2. Hierarchical Processing:

   • Lower layers focus on local patterns (conv/recurrent)
   • Upper layers capture global dependencies (attention)

3. Memory Efficiency:

   • Streaming tokenization during vocab building
   • No full dataset storage in RAM
   • Automatic cleanup of intermediate data

Model Files

• pytorch_model.bin: Model weights
• config.json: Model configuration
• tokenizer.json: Tokenizer vocabulary

Limitations

• Trained on English text only
• Limited to 256 token context window
• May require fine-tuning for specific downstream tasks
• Conversational style influenced by TinyChat dataset

Model Series

• i3-22M - Original model with pure hybrid architecture
• i3-80M - Scaled version with attention layers and multi-dataset training
• i3-200M (This model)

Citation

@article{mamba,
  title={Mamba: Linear-Time Sequence Modeling with Selective State Spaces},
  author={Gu, Albert and Dao, Tri},
  journal={arXiv preprint arXiv:2312.00752},
  year={2023}
}
@article{RWKV,
  title={RWKV: Reinventing RNNs for the Transformer Era},
  author={Peng, Bo and others},
  journal={arXiv preprint arXiv:2305.13048},
  year={2023}
}