Tag: #finetuning6 results
Contrastive Decoding Diffing (CDD) is a new grey-box method that recovers verbatim training data from finetuned LLMs using only logit access. It outperforms existing white-box methods like Activation Difference Lens (ADL) without requiring weight access or complex calibration.
User runs Qwen3 TTS locally in real-time for expressive VTuber pipeline. Implemented streaming, llama.cpp support, CTC alignment, and finetuned a custom voice. Shares GitHub repo for Persona Engine.
Hugging Face blog details training and finetuning multimodal embedding and reranker models using Sentence Transformers. It covers handling text and images for advanced retrieval tasks. Developers gain open-source tools for building superior RAG systems.
New synthetic finetuning dataset from Claude Opus 4.6/4.7 with 8,706 reasoning examples across 28 categories. Includes splits for instruct, roleplay, and code; totals 17M tokens. Hosted on Hugging Face with detailed stats.
A newbie finetuned Llama 3.1 3B-Instruct on WhatsApp chats to mimic texting style, but the model references unrelated past events like math tests. It responds to simple greetings with off-topic queries. User seeks advice to fix this behavior.
User finetuned Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-GGUF for $3 in 10 minutes using original dataset. Result: cleaner reasoning, less bloat, maintained or improved accuracy vs. problematic original. Highlights easy finetuning potential for beginners.