Qadir v1 (v2 in progress)
Qadir is the agent brain of QADIR_CORE. It is a set of LoRA adapters for Qwen/Qwen2.5-7B-Instruct-1M (Apache-2.0), trained to call tools and keep a consistent persona.
Files
| Path | What it is |
|---|---|
adapters/qadir_v1_tools/ |
Tool-use LoRA, r=64 (the main adapter) |
adapters/qadir_v1_personality/ |
Persona LoRA (the "clean" run), r=32 |
evals/qadir_v1_eval.json |
Raw eval |
evals/qadir_v1_eval_upgraded.json |
Eval with 2-shot examples and forced grammar on missed tools |
Honest scores (480 tool prompts, 48 tools, Q8 GGUF)
| Setup | Accuracy | Tools at the 0.9 gate |
|---|---|---|
| Raw | 71.67% (344/480) | 21/48 |
| With prompting (2-shot + forced grammar) | 88.75% (426/480) | 37/48 |
v1 does not pass its own 0.9 gate. It is published as a working baseline, not a finished product.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct-1M", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "abuz8studios/Qadir-v1-v2", subfolder="adapters/qadir_v1_tools")
tok = AutoTokenizer.from_pretrained("abuz8studios/Qadir-v1-v2", subfolder="adapters/qadir_v1_tools")
Roadmap
- v2: in progress, no eval yet. The humility and tool-retraining runs exist locally but are not graded, so they are not published.
- Next steps: pass the 0.9 gate on all 48 tools without prompt tricks, then publish GGUF builds.
About the lab
ABUZ8 is a one-person lab. Ahmad builds it alone, with no team and no funding, and has spent the last year living inside AI work. The long-term goal is the SUPERBRAIN fusion: Arabic-root reasoning (SUPERBRAIN), tool use and memory (Qadir), fast local inference (FAIE) and new architecture experiments (MUTHALLATH), all joined in one open system.
- Downloads last month
- -