| Benchmark | Base webe tori | 0105 Patched | Improvement | |-----------|----------------|--------------|--------------| | EQ-Bench (instruction following) | 42.3 | 68.7 | +26.4 pts | | Repetition (500 tokens, temp=1.0) | 14% loop | 2% loop | 12% better | | Coherence (1-10 score) | 6.2 | 8.5 | +37% | | Multi-turn consistency (4 turns) | 31% drift | 8% drift | 23% better | Note: These are community-aggregated estimates, not official results from a paper. If you’ve found a copy of this patched model (e.g., on Hugging Face under a user like webe/tori-0105-patched or via a Torrent/AI mirror), here’s how to run it effectively: 1. With llama.cpp (GGUF version) ./main -m webe-tori-0105-patched.Q4_K_M.gguf -n 512 -p "User: Write a haiku about patched AI. Assistant:" -temp 0.8 -repeat_penalty 1.12 2. With Transformers (PyTorch) from transformers import AutoModelForCausalLM, AutoTokenizer model_name = "webe/tori-0105-patched" # Example path tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
Next time you encounter a broken model on Hugging Face, remember the tale of webe tori. With a little effort and the right patch, even a flawed bird can learn to fly straight. Have you used the webe tori model 0105 patched? Share your experience in the comments below or contribute your own patch findings to the community. webe tori model 0105 patched
| Issue | Description | |-------|-------------| | | Random <0x09> or </s> tokens appearing mid-generation. | | Repetition penalty mismatch | The model ignored repetition penalties, leading to loops after 200 tokens. | | Instruction drift | After 3 conversational turns, the model reverted to base-model behavior (e.g., acting like a generic assistant). | | Sampling instability | High temperature (1.1+) caused gibberish output more than expected. | | Benchmark | Base webe tori | 0105