Llama 3.1 70B TFree-HAT SFT
The 70B TFree-HAT SFT release adapts Llama 3.1 70B's pretrained backbone with character-level encoding and decoding. It is further trained in English and German and released for research and education.
Research release
Model design
The hierarchical architecture aims to compress text into fewer backbone positions. The model card separates compression results from realized inference throughput and notes that the public implementation is still being optimized. It is not a code- or mathematics-specialist release.Llama 3.1 70B TFree-HAT SFT model card
Practical use
Practical guidance. Evaluate the large checkpoint only when its additional language capacity matters to the research question. Account for both the backbone and custom character processing in serving measurements. Do not translate a text-compression gain directly into a hardware-speed claim.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | Aleph-Alpha/llama-3_1-70b-tfree-hat-sft |
| License | Open Aleph License for non-commercial research and education.Llama 3.1 70B TFree-HAT SFT model card |
| Ancestry | Llama 3.1 70B Base.Llama 3.1 70B TFree-HAT SFT model card |
| Serving | Model-specific inference support is required.Llama 3.1 70B TFree-HAT SFT model card |
Related models
- Llama 3.1 8B TFree-HAT SFT
Aleph Alpha ยท TFree-HAT
Tokenizer-free Llama 3.1 8B adaptation with supervised instruction training in English and German.
Launch date not confirmedResearch releaseDownloadable