Downloads · 30 days
0
qkz-opus4.7-aieng-3b-v2.1-adapter
qukaizen/qkz-opus4.7-aieng-3b-v2.1-adapter
qkz-opus4.7-aieng-3b-v2.1-adapter is a machine learning model from qukaizen. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
- adapters.safetensors — trained LoRA weights (8.5 MB) - adapterconfig.json — training configuration (mlxlm format)
Access
Public
Updated May 18, 2026
Repo size
8.9 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors8.9 MB · 100%
At a glance
- Access
- Public
- Created
- May 18, 2026
- Updated
- May 18, 2026
- SHA
- 963a078a
Model card
From the Hugging Face model README
LoRA Adapter — qkz-opus4-aieng-3b-v2.1
Files
adapters.safetensors— trained LoRA weights (8.5 MB)adapter_config.json— training configuration (mlx_lm format)
Base model
mlx-community/Qwen2.5-3B-Instruct-4bit (Qwen2.5-3B-Instruct, 4-bit quantized for MLX).
Note: the parent superskill-spec.yaml lists student_base: phi-3.5-mini-instruct
because that's what the manifest declared. The trainer was launched with
STUDENT_MODEL=mlx-community/Qwen2.5-3B-Instruct-4bit (see adapter_config.json),
so the actual weights you're loading were trained against Qwen2.5-3B. Treat
adapter_config.json as authoritative.
LoRA parameters
| Rank | 16 |
| Scale (alpha/rank) | 1.0 |
| Dropout | 0.05 |
| Target modules | self_attn.q_proj, self_attn.v_proj, self_attn.k_proj |
| Trainable layers | 16 (top of model) |
| Trainable params | 2.228M / 3.086B (~0.072%) |
Training config
| Iterations | 90 (3 cycles × ~30 iters/cycle) |
| Batch size | 1 |
| Gradient accumulation | 16 steps |
| Learning rate | 2e-4 |
| Max seq length | 4096 |
| Optimizer | Adam |
| Seed | 42 |
Load with mlx_lm
from mlx_lm import load, generate
model, tokenizer = load(
"mlx-community/Qwen2.5-3B-Instruct-4bit",
adapter_path="./adapter", # this directory
)
print(generate(model, tokenizer, prompt="Explain LoRA fine-tuning.", max_tokens=256))
Load with the trainer's /infer endpoint
curl -X POST http://localhost:8006/infer \
-H "Authorization: Bearer $NUCLEUS_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt":"Explain LoRA fine-tuning.","adapter_path":"/absolute/path/to/this/adapter","max_tokens":256}'
- Created
- May 18, 2026
- Updated
- May 18, 2026