JewGPT-Qwen3.5-9B

JewGPT-Qwen3.5-9B is a fused MLX checkpoint derived from mlx-community/Qwen3.5-9B-BF16. It was produced by fine-tuning a LoRA adapter on a small custom chat dataset and then merging that adapter back into the base model weights.

Overview

• Base model: mlx-community/Qwen3.5-9B-BF16
• Fine-tuning method: LoRA
• Fused checkpoint: fused_model_0001000
• Intended task: text generation / chat-style responses

Training Details

The local training configuration in this project indicates:

• Training dataset: jewgpt_dataset/train.jsonl
• Dataset size: 102 chat examples
• Iterations: 1000
• Learning rate: 1e-5
• Batch size: 1
• Gradient accumulation steps: 1
• Max sequence length: 2048
• LoRA rank: 8
• LoRA dropout: 0.0
• LoRA scale: 20.0
• Targeted layers: 16
• Seed: 0

Behavior Notes

This model is highly specialized and reflects the narrow tone, framing, and persona present in its training examples. It should be treated as an experimental derivative rather than a general purpose assistant.

Because the fine-tuning dataset is small and strongly opinionated, outputs may be:

• roleplay-heavy
• politically biased
• overconfident
• unreliable for factual or safety-critical use

Limitations

• Not evaluated against standard benchmarks
• Not suitable as a source of factual political analysis
• Not suitable for high-stakes use
• Likely to overfit the training persona and response style

Files

This repository contains the merged model shards, tokenizer files, and chat template needed to run the fused checkpoint.

License

This release is a derivative of Qwen3.5-9B and follows the upstream licensing information linked above. Please review the base model license and ensure your use complies with it.