AICE-v1 Model Card

Overview

AICE-v1 is a compact conversational model in RWKV architecture, distributed as merged weights for text generation. AICE-v1 is a derivative model initialized from externally pre-trained foundation weights, then innovated with LoRA and merged for inference.

Technical profile

• Architecture: RWKV causal language model
• Layers: 24
• Hidden size: 2048
• Context length: 1024
• Vocabulary size: 50277
• Effective size: about 1.5B parameters
• Primary format: model.safetensors (FP16/F32 mixed tensors, runtime friendly)

Training strategy

• Initialization from pre-trained foundation checkpoint weights (third-party origin).
• Instruction tuning performed with LoRA adapters.
• Distilled supervision pipeline derived from a larger teacher model family (70B class).
• LoRA adapters merged into a single consolidated model for inference.

Release intent

• This repository contains the merged model artifacts for inference.
• Adapter artifacts are optional internal training artifacts and are not required for runtime.

Suggested use

• Assistant/chat inference
• Lightweight deployment scenarios (desktop and mobile with quantized variants)
• Prompt-based reasoning tasks

Limitations

• Behavior quality depends on prompt design and decoding setup.
• The model can still produce hallucinations and incorrect factual outputs.
• Safety filtering and evaluation are required in production.

Runtime files

• config.json
• generation_config.json
• tokenizer.json
• tokenizer_config.json
• special_tokens_map.json
• model.safetensors

Mobile quantization

See MOBILE_Q4_PIPELINE.md for INT8/INT4 export and packaging.

Formats note

• Included now: safetensors, ONNX (INT8, INT4) and GGUF.
• GGUF artifact: aicemobile/AICE_v1_rwkv4_custom.gguf (custom RWKV4 GGUF layout).

Compliance documentation

• EU AI Act public training-content summary: EU_TRAINING_SUMMARY.md