kth8/gemma-3-270m-it-JSON-Fixer-GGUF

A fine-tune of unsloth/gemma-3-270m-it on the kth8/json-fix-25000x dataset.

Usage example

System prompt

You are a JSON formatting specialist. Convert the provided JSON data into valid JSON format with 2 line indent and no additional commentary.

User prompt

The JSON is:\n[{\"name\":\"John Doe\", \"jobTitle\":Software Engineer, \"department\": \"Research and Development\"],, {\"name\"\"Jane Smith\", \"jobTitle\":\"Data Analyst', \"department\":\"Marketing and Sales\"}, ]    //\" comment\n-- end --

Assistant response

[
  {
    "name": "John Doe",
    "jobTitle": "Software Engineer",
    "department": "Research and Development"
  },
  {
    "name": "Jane Smith",
    "jobTitle": "Data Analyst",
    "department": "Marketing and Sales"
  }
]

Model Details

• Base Model: unsloth/gemma-3-270m-it
• Parameter Count: 268,098,176
• Precision: torch.bfloat16

Hardware

• GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
• Announced: Mar 17th, 2025
• Release Date: Mar 18th, 2025
• Memory Type: GDDR7
• Bandwidth: 1.79 TB/s
• Memory Size: 96 GB
• Memory Bus: 512 bit
• Shading Units: 24064
• TDP: 600W
• FP16 (half): 126.0 TFLOPS (1:1)

Training Settings

PEFT

• Rank: 32
• LoRA alpha: 64
• Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Gradient checkpointing: unsloth

SFT

• Epoch: 2
• Batch size: 32
• Gradient Accumulation steps: 1
• Warmup ratio: 0.05
• Learning rate: 0.0004
• Optimizer: adamw_torch_fused
• Learning rate scheduler: cosine
• Max seq length: 2048

Training stats

• Date: 2026-03-23T04:39:38.019077
• Peak VRAM usage: 64.5 GB
• Global step: 1538
• Training runtime (seconds): 1142.9274
• Average training loss: 0.004019292104312295
• Final validation loss: 0.0014343492221087217

Framework versions

• Unsloth: 2026.3.10
• TRL: 0.22.2
• Transformers: 4.56.2
• Pytorch: 2.10.0+cu128
• Datasets: 4.8.3
• Tokenizers: 0.22.2

License

This model is released under the Gemma license. See the Gemma Terms of Use and Prohibited Use Policy regarding the use of Gemma-generated content.