Fine-tuning

Nous released Qwen-based Arete-27B, training data and an 803-prompt evaluation. It reports cutting its measured propaganda rate from 22% to 1% without a significant Hermes Bench drop.

Thinking Machines released Inkling 2 Small on Tinker with native vision and fine-tuning support. Apache-licensed weights are planned after safety testing, while early checkpoints required fixes for guardrail circumvention.

Transluce released DSSE-1.1, containing 6,915 datasets across 18 unexpected fine-tuning behaviors. It reports that side effects transfer across models, while researchers predicted outcomes with 63% accuracy against 50% chance.

FineDNA released 514B mid-training tokens, 200,000 SFT conversations and 100,000 RL tasks. Its authors report about 67% variant-classification accuracy with code-grounded explanations, versus below 60% for self-distillation.

New LFM2.5-24B LoRA adapters target Pi and Claude Code, with GGUF weights available. The author reports up to 40% fewer Pi tool calls, but no meaningful improvement for the Claude Code variant.

Mistral released Large 4 weights with launch-day serving and LoRA training support. vLLM reports roughly twice baseline output throughput with its built-in drafter in specified tests.

Kev 1.0 introduces Kev-27B and updates Kev-9B with a 64k document window and TypeSafe SDK compatibility. The release includes weights, source code, and fine-tuning and deployment skills.

Experiments reported by Maxime Rivest found task-specific small classifiers outperforming frontier models on specialized decisions. One result says a 4-million-parameter BERT Tiny model beat Opus and Kimi after training on 10,000 examples.

LangSmith Fine-Tuning and the open-source smithtune CLI now turn agent traces into fine-tuning datasets. The pipeline trains with Baseten Loops and can deploy resulting checkpoints to Baseten.