doobz1111/idiotbot-v3-gguf

IdiotBot v3 (Q4_K_M GGUF)

A fine-tuned Qwen3-8B model trained on Australian IRC channel logs. Captures the casual, irreverent tone of Australian internet culture — witty, crude, and unapologetically Aussie.

Model Details

Base Model	Qwen/Qwen3-8B (Instruct)
Method	QLoRA (4-bit) via Unsloth
LoRA Rank	16 (alpha 32)
Training Data	~90K curated multi-turn IRC conversations
Format	ChatML
Sequence Length	1024
Epochs	1
Learning Rate	2e-5 (cosine schedule)
Effective Batch Size	16 (8 x 2 gradient accumulation)
Quantization	Q4_K_M (4.7 GB)
Hardware	NVIDIA H100 NVL (~2.5 hours)

Training Results

Step	Train Loss	Eval Loss
2000	1.52	1.50
3000	1.43	-
4000	1.40	-
5000	1.39	1.38

Train and eval loss converged closely, indicating good generalisation with no overfitting.

Training Data

Sourced from IRC chat logs spanning multiple years across several Australian channels. Over 1 million raw messages were cleaned, threaded into conversations, and quality-scored, resulting in ~90K high-quality multi-turn training examples.

Data Pipeline

1. Parse IRC logs to structured JSONL
2. Clean noise, normalise nicks, thread conversations
3. Pattern-based annotation (topics, quality scoring, tone)
4. Quality-weighted selection (higher quality conversations upweighted)
5. Format as ChatML with channel context

Usage with Ollama

ollama create idiotbot -f Modelfile

Example Modelfile:

FROM ./idiotbot-v3-q4_k_m.gguf

TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
{{ .Response }}<|im_end|>"""

PARAMETER temperature 0.7
PARAMETER top_p 0.8
PARAMETER top_k 20
PARAMETER repeat_penalty 1.1
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"

Recommended Inference Settings

Parameter	Value
temperature	0.7
top_p	0.8
top_k	20
repeat_penalty	1.1
think	false

Set think: false to disable Qwen3 thinking mode.

Limitations

• Trained specifically on Australian IRC chat — not a general-purpose model
• Contains casual language, slang, and crude humour
• Not suitable for professional or sensitive applications
• Designed as an IRC chatbot persona, not an assistant

License

Apache 2.0 (inherits from Qwen3-8B base model)