naksyu/gemma4-e4b-diagnostic-v2-fft100

VISIONConcurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 9, 2026Architecture:Transformer Featherless Exclusive Cold

The naksyu/gemma4-e4b-diagnostic-v2-fft100 is an experimental 7.94 billion parameter Gemma4 E4B full-finetuned model, based on the Gemma4ForConditionalGeneration architecture. It is specifically designed for short training-log diagnosis, metric interpretation, checkpoint triage, and providing compact reasoning answers. This model excels at identifying training failure modes and making concrete next-step decisions from numerical logs, primarily in English and some Korean.

Loading preview...

Overview

naksyu/gemma4-e4b-diagnostic-v2-fft100 is an experimental 7.94 billion parameter model, full-finetuned from a local gemma4-e4b-it checkpoint. It utilizes the Gemma4ForConditionalGeneration architecture and is published in BF16 model.safetensors format. This model is a research artifact, not intended for production deployment without further validation, and focuses on a very specific niche: training diagnostics.

Key Capabilities

  • Training Log Diagnosis: Designed to interpret training logs and metrics.
  • Failure Mode Identification: Can identify issues like plateauing, overfitting, gradient norm problems, OOM errors, and label coverage problems.
  • Checkpoint Triage: Assists in selecting better checkpoints based on evaluation metrics.
  • Concise Reasoning: Provides short, visible rationales and concrete decisions based on numerical data.
  • Language Support: Primarily covers English with some Korean.

Training Details

The model was trained using a full fine-tune method, not LoRA, with a maximum sequence length of 1024. It was trained for 300 steps with saves every 100 steps, making this a snapshot from an intermediate experimental run. The training data, totaling 30,156 rows (28,156 train, 2,000 eval), was specifically curated from sources like OpenThoughts-114k and synthetic diagnostic datasets to teach short rationale and concrete training decisions. It explicitly excluded previous data/sft mixes, persona/style sources, tool-call rows, and hidden-thought tags.

Intended Use Cases

This model is specifically intended for:

  • Interpreting eval_loss: nan as mechanical rather than model collapse.
  • Summarizing trends without verbose hidden-thought style output.
  • Assisting in making concrete decisions regarding model training progression.

Limitations

  • Lacks a full benchmark report.
  • May hallucinate training causes or overstate confidence.
  • Not a replacement for actual evaluation scripts or dataset inspection.
  • Tuned as text-only, despite the base architecture potentially supporting multimodal tokens.
  • Safety behavior was not a primary target of this fine-tune.