semantic-sft checkpoints

Research checkpoints of the semantic-sft project: how SFT injects new facts (fictional people's biographies) into an instruction-tuned model, and what it costs the model. Research only; the facts are fictional and the models are degraded on purpose in many runs.

  • One folder per run: <run>/e<epoch>/ (config.json, generation_config.json, model.safetensors, bf16) and <run>/run-config.json (all settings and data hashes). Load: AutoModelForCausalLM.from_pretrained("kimgreg/semantic-sft", subfolder="<run>/e20"); the tokenizer is the base model's.
  • Modified files: every checkpoint is a fine-tuned modification of LFM2.5-230M (Liquid AI), distributed under the LFM Open License v1.0 (LICENSE, a copy of the base model's license).
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for kimgreg/semantic-sft

Finetuned
(50)
this model