AI Briefing#sft
The latest AI and developer news about #sft, with the original source and a short summary.
The latest AI and developer news about #sft, with the original source and a short summary.
SFT on 52.8 million tokens reduced the Chinese government narrative adoption rate from 22% to 0.9%.
A three-week engagement used near-policy SFT and SDPO to lift the model's quality by 20.7 points over the base and 8 points over the frontier competitor.
Second Look Research demonstrates that combining fictional story SDF and difficult advice SFT reduces agentic misalignment in Qwen 2.5 32B and 72B models to 4%, with difficult advice SFT alone reducing it to 6% and significantly outperforming stories alone.
A rank-16 LoRA distilled from representation analysis cuts the slop score by 83% without system prompts or significant quality loss.
Olive Young unveiled an alignment learning approach that simultaneously improves the naturalness and format compliance of AI Pick recommendation cards by leveraging NLL metrics.
A Reasoning Corpus dataset with 5 million entries designed to improve the reasoning performance of small language models (SLMs) has been released.
SupraLabs has released a 50M-scale reasoning model that generates a thinking process using a synthetic dataset.
By applying DPO to OCR tasks, text degeneration that is hard to resolve with SFT alone was significantly improved.
This covers a case where Anthropic improved Claude's ability to respond to blackmail through a small amount of high-quality data.
The author compares SFT, RL, and OPD from a distributional perspective and emphasizes the importance of policy-based sampling.
Qwen 3.6 27B was remodeled with Opus CoT and Hermes tool-calling SFT and released in FP8.
Covers directions for improving the performance of language models combined with search.
BAR is an approach where domain-specific experts are each post-trained independently and then merged into an MoE.
By specializing 3B and 7B open-source SLMs with SFT+DPO, they achieved higher scores than commercial models.
Netflix built a flexible LLM Post-Training Framework on top of Ray, PyTorch, and vLLM.
Olive Young optimized its review theme recommendation system by building a Gemma 3-based sLLM in a constrained environment with just one T4 GPU.
Agentic capabilities were enhanced without degrading Korean language performance by applying mid-training and a parallel RL pipeline.
SyGra, a low-code framework supporting data generation, transformation, and alignment for LLM and SLM training, has been released.
AIME benchmark performance was improved through high learning rates in the SFT stage and relaxed clipping strategies in RL.
The Kanana LLM, developed through SFT, DPO, and model merging, demonstrated superior performance compared to similar-sized models on Korean benchmarks such as LogicKor.