Yoro9381/LFM2.5-1.2B-Instruct-Korean-Opus-4.6-Distill

TEXT GENERATIONConcurrency Cost:1Model Size:1.2BQuant:BF16Ctx Length:32kPublished:Jul 1, 2026License:lfm-open-license-v1.0Architecture:Transformer Open Weights Cold

Yoro9381/LFM2.5-1.2B-Instruct-Korean-Opus-4.6-Distill is a 1.2 billion parameter instruction-following language model, fine-tuned from LiquidAI/LFM2.5-1.2B-Instruct. This model specializes in Korean language tasks, including question answering, general conversation, and instruction following, with a notable context length of 32768 tokens. It was trained on Korean-centric datasets to generate natural, consistent, and contextually appropriate responses in Korean. The model is optimized for stable generalization performance in Korean language applications.

Loading preview...

Model Overview

Yoro9381/LFM2.5-1.2B-Instruct-Korean-Opus-4.6-Distill is a 1.2 billion parameter instruction-following language model, building upon the LiquidAI/LFM2.5-1.2B-Instruct base model. It has been specifically fine-tuned to enhance its performance in Korean language tasks, leveraging a substantial context length of 32768 tokens.

Key Capabilities & Training

This model excels in:

  • Korean Instruction Following: Designed to accurately understand and respond to instructions in Korean.
  • Korean Question Answering: Provides relevant and contextually appropriate answers to Korean queries.
  • General Korean Conversation: Capable of engaging in natural and consistent dialogue in Korean.

Training involved three full epochs over 2,160 steps, utilizing diverse Korean-centered datasets such as maywell/koVast, CarrotAI/ko-instruction-dataset, MarkrAI/KoCommercial-Dataset, and Jongsim/claude-opus-4.6-reasoning-12k-ko-filtered-v2. This broad dataset exposure aimed to capture a wide range of Korean expressions and contexts.

Performance & Stability

Evaluation metrics showed a training loss of 0.9999 and an evaluation loss of 1.0480, indicating stable generalization performance without significant overfitting. The eval_mean_token_accuracy of 0.7445 suggests consistent next-token prediction. This stability implies the model effectively learned major patterns in the training data.

Limitations

While training metrics are strong, the model's real-world usability, factuality, and response quality require further qualitative and quantitative assessments, including user question-answer evaluations and safety checks. The model's use is governed by the LFM Open License v1.0 of its base model.