eigentiki

I trained this LoRA adapter on top of google/gemma-2-9b-it. I used 6,625 examples of real coding agent work (tool calls, bash, reasoning).

Training

• Method: QLoRA (4-bit) with Unsloth
• Data: focustiki/sft-coding-agent-traces
• I ran a sweep of 3 configs and picked the best by eval loss.
• Winner: run_1 (eval_loss=4.5200)
• Settings: lr=0.0002, LoRA r=16, seq_len=4096

Sweep Results

Run	LR	LoRA r	Seq Len	Eval Loss	Train Loss
run_1	0.0002	16	4096	4.5200	5.6922
run_2	0.0001	32	8192	4.6942	5.9130
run_3	5e-05	64	8192	4.8814	6.1487

How to Use

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="focustiki/eigentiki",
    max_seq_length=8192,
    load_in_4bit=True,
)