Base GPT-2 XL Better SFT โ€” Hard Repair Continuation

Continuation checkpoint trained from Darkstorm1826/base-gpt2xl-better-sft using a JAX/Optax TPU full fine-tune pass.

This pass directly targets:

  • fake current facts
  • fake citations and fake URLs
  • invented context causes
  • repeated filler and output degeneration
  • unrelated metadata/story continuations
  • greeting/script contamination
  • nationality/self-introduction contamination
  • Nintendo/Mario factual QA errors
  • short context-faithful answers
  • basic arithmetic patterning

Prompt format:

<|system|> You are a helpful, honest, and careful AI assistant.<|end|> <|user|> Question here<|end|> <|assistant|> Answer here<|end|>

Downloads last month
100
Safetensors
Model size
2B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Darkstorm1826/base-gpt2xl-better-sft

Unable to build the model tree, the base model loops to the model itself. Learn more.