awilliamson/talkie-1930-13b-it-vllm
awilliamson/talkie-1930-13b-it-vllm is a 13 billion parameter instruction-tuned language model, repackaged for vLLM and HuggingFace compatibility. Developed by talkie-lm, its base model was pretrained on approximately 260 billion tokens of pre-1931 English text, with instruction tuning on pre-1931 reference works. This model specializes in generating text with a historical linguistic style, making it suitable for tasks requiring an authentic pre-1930s English voice. It features a 32768 token context length and is optimized for efficient serving with vLLM.
Loading preview...
Model Overview
awilliamson/talkie-1930-13b-it-vllm is a 13 billion parameter instruction-tuned model, originally from the talkie-lm project, now repackaged for seamless integration with HuggingFace transformers and efficient serving via vLLM. The base model was extensively pretrained on approximately 260 billion tokens of English text dating before 1931. Its instruction-tuned variant was refined using a dataset derived from pre-1931 reference materials, such as etiquette manuals, encyclopedias, and letter-writing guides, further optimized with online DPO.
Key Capabilities
- Historical Text Generation: Excels at producing text that authentically reflects the linguistic style and nuances of pre-1930s English.
- vLLM Compatibility: Includes necessary
config.json,tokenizer.json, and custom modeling code (modeling_talkie.py,configuration_talkie.py) to enable out-of-the-box vLLM serving. - HuggingFace Integration: Fully compatible with HuggingFace
AutoTokenizerandAutoModelForCausalLMfor standard inference workflows. - Instruction Following: Instruction-tuned on historical reference works, allowing it to respond to prompts in a period-appropriate manner.
Good For
- Historical Research & Simulation: Generating content for historical fiction, academic research, or simulating conversations from the early 20th century.
- Creative Writing: Crafting narratives, dialogues, or documents that require an authentic pre-1930s English voice.
- Specialized Applications: Any use case where a model's output needs to adhere to the linguistic conventions of a specific historical period.
Note on Sampling: For optimal results, it is recommended to use temperature >= 0.5 during sampling, as greedy decoding (temperature=0) can lead to repetitive outputs.