Fred-9B

Fred-9B is a fine-tuned 9B-parameter language model based on Qwen 3.5 9B and finetunned with this dataset, https://huggingface.co/datasets/CrowdMind/caveman-fred.

Fred-9B was trained with LoRA/QLoRA-style parameter-efficient fine-tuning and is intended as an experimental conversational and reasoning model.

Note: Fred-9B is a fine-tune of Qwen 3.5 9B, not a model trained from scratch.

Model Details

Property Value

---

Model name Fred-9B Base model Qwen 3.5 9B Fine-tuning method LoRA LoRA rank 128 LoRA alpha 256 LoRA dropout 0.05 Batch size 4 Context length 2,048 tokens Learning rate 2e-4 Optimizer AdamW 8-bit Warmup steps 30 Maximum training steps 300 Training tokens ~3.21M Final training loss 0.0833 Final gradient norm 0.470 Training time ~1h 33m

Training

Fred-9B was fine-tuned for 300 steps. The training run used a 2,048-token context length and a batch size of 4.

The learning-rate schedule used a short warmup followed by decay across the training run.

Final Training Metrics

• Training loss: 0.0833
• Learning rate: 7.41e-7
• Gradient norm: 0.470
• Tokens processed: 3,214,912
• Steps: 300 / 300
• Training time: 1h 32m 45s

No evaluation dataset was configured for this training run, so an evaluation loss was not recorded.

Example Behavior

One of the training checkpoints was tested with simple arithmetic and order-of-operations questions.

Example 1

Prompt

Please solve this for me i need help: 2+1*3(1+1)+3

Fred-9B

The final answer is 11.

The model correctly handles the parentheses and multiplication before addition:

1 + 1 = 2
3 × 2 = 6
2 + 6 + 3 = 11

Example 2

Prompt

Please solve this for me i need help: 3+1+1*4(1+1)+3*1

Fred-9B

The final answer is 15.

The model breaks the expression down using standard order of operations and reaches the correct result.

Intended Use

Fred-9B is intended for:

• Conversational AI experiments
• Local LLM experimentation
• Coding and reasoning experiments
• Mathematical/problem-solving prompts
• Testing fine-tuning workflows
• Personal/local AI assistants

Because Fred-9B is an experimental fine-tune, its behavior may vary substantially depending on the prompt and task.

Limitations

Fred-9B has several important limitations:

• The model has only been fine-tuned for a relatively small number of steps.
• The reported run did not use an evaluation dataset.
• The examples shown above are not a comprehensive benchmark.
• Correct answers on simple arithmetic do not guarantee reliable mathematical reasoning.
• The model may hallucinate or produce incorrect information.
• Fine-tuning can alter some of the capabilities and behaviors of the original Qwen model.
• Performance will depend on the inference engine, quantization, prompt format, and hardware.

For serious evaluation, use a held-out evaluation dataset and established benchmarks rather than relying only on training loss.

Recommended Evaluation

For future versions of Fred, consider evaluating:

• General instruction following
• Mathematical reasoning
• Coding
• Knowledge/retrieval
• Multi-step reasoning
• Hallucination resistance
• Long-context behavior
• Conversational consistency

A separate validation/evaluation dataset should be configured for future training runs so that training loss can be compared against evaluation loss.

Running Fred-9B

Fred-9B can be used with an inference stack that supports the model's exported format.

If the model is exported to GGUF, it can be used with compatible local inference tools such as llama.cpp-based applications.

Example concept:

# Example only — replace with the actual exported model filename
./llama-cli -m Fred-9B.gguf

The exact command depends on the exported model format and inference framework.

Quantization

For local deployment, quantized versions can significantly reduce memory requirements.

Possible variants include:

• 4-bit
• 5-bit
• 6-bit
• 8-bit
• Full/half precision

The best choice depends on available VRAM/RAM and the desired balance between speed, memory usage, and model quality.

Training Configuration

The training run used approximately the following configuration:

model: Qwen 3.5 9B

method: LoRA

training:
  batch_size: 4
  learning_rate: 0.0002
  optimizer: AdamW 8-bit
  context_length: 2048
  warmup_steps: 30
  max_steps: 300

lora:
  rank: 128
  alpha: 256
  dropout: 0.05
  variant: lora

Acknowledgements

Fred-9B is based on Qwen 3.5 9B.

All original model licensing, attribution, usage restrictions, and terms from the Qwen base model remain important when using or redistributing Fred-9B. Check the official Qwen model license and documentation before publishing or deploying the model.

Project Status

Experimental --- first training run

Fred-9B is currently an experimental fine-tune. Future versions can improve the dataset, evaluation process, training duration, instruction formatting, and inference quality.

---

Fred-9B --- a personal 9B fine-tune built on Qwen 3.5 9B.