media Hugging Face Forums · 1h ago · open_models

Eunoia-Logos-Adri open-sources The Forge full fine-tuning pipeline for Eunoia

from English

The group behind the Eunoia model, built on a Llama-3.1-Dolphin-3.0-8B base, has open-sourced "The Forge," a specialized architectural pipeline and algorithmic fine-tuning strategy under the CC0 1.0 Universal license.

The approach prioritizes structural crystallization over unbounded context windows to avoid memory saturation and informational noise.
The raw code includes optimizations for bare-metal CPU clusters, such as disabling gradient clipping and using native FP32 caching to prevent RAM fragmentation.
A custom dataset class enforces a strict linear sequence order to preserve the contextual structure of historical data during training.
The release includes the AGIO-CIMIENTO-dataset.txt, over 500,000 words of evaluation tracking logs, and raw GPQA Diamond evaluation results to verify general capability retention.

The authors publish the experiment from its start to encourage independent reproduction and testing on faster hardware, noting that the final mathematical results will only be known after several days of training.

Importance 1/3 New harness with differentiators Trust 1/3 Hugging Face Forums Open weights Training methods

Related articles

media MarkTechPost · 22d ago · 26 views

Google Research releases ToolGrad, a tool-use data generation framework

Researchers from Google and several Japanese universities have introduced ToolGrad, a framework that inverts the traditional pipeline for generating tool-use datasets by constructing verified API chains first and then writing matching user queries. This approach aims to eliminate the inefficiency of query-first methods that often fail during agent exploration.

media MarkTechPost · 26d ago · 47 views

IFM releases K2 Horizon: six Apache 2.0 models from 0.9B to 375B

The Institute of Foundation Models (IFM) released K2 Horizon, a fleet of six open-source large language models ranging from 0.9B to 375B parameters, all licensed under Apache 2.0.

arxiv arXiv cs.CL · 32d ago · 50 views

CogEvol releases 4B model for efficient learning environment generation

CogEvol introduces a family of models designed to convert course briefs into structured JSON slides or interactive HTML pages in a single pass, replacing multi-turn agent scaffolding. The system utilizes a production-grounded data pipeline and hybrid reward mechanisms to enforce reliability during training.

arxiv arXiv cs.AI · 37d ago · 44 views

Thomson enables SovereignAI via continual learning on open-weight models

The article introduces Thomson, a general-purpose frontier model developed by SovereignAI that demonstrates how institutions can achieve frontier performance through continual learning on readily available open-weight models. This approach utilizes a modern mid- and post-training stack with safeguards to preserve plasticity and stability, allowing for minimal high-impact interventions on parameters.

arxiv arXiv cs.AI · 52d ago · 36 views

MiLMMT-46-v1.0 improves multilingual translation via reference-free post-training

Researchers have developed MiLMMT-46-v1.0, a multilingual machine translation model that applies reference-free post-training to open large language models. The team uses Group Relative Policy Optimization (GRPO) with a reward based on two reference-free quality estimation models, gated by language identification.