INTENT RLSE Facts 500
Facts fine-tuning checkpoint at step 500. Parent: SFT-v2 step 750; pretraining step 10000. Custom RLSE architecture, 512-token trained context. This is a PyTorch state dictionary, not a Transformers AutoModel package.
Compared against deployed SFT 1000: validation perplexity 200.64 → 9.33 and teacher-forced end-token accuracy 0% → 100%. Factual generation remains weak: only 1/15 exact matches with greedy decoding. See comparison and raw results for methods, limitations and outputs.
Run
Download the files, install requirements, then:
python infer.py --checkpoint ckpt-facts-00500.pt --prompt "What is the capital of France?" --temperature 0
Defaults inherited from the requested runtime are temperature 9 and top-k 80; temperature 9 yielded no exact matches in the 15-question probe. No reliable factual accuracy is claimed. The model may confidently output incorrect facts.
The checkpoint was uploaded separately because the existing demo Space cannot hold both weights under its 1 GB limit. Uploading does not change its active model. No license grant is asserted here; training provenance is the supplied facts branch based on TriviaQA and custom factual examples.