Hugging Face's logo

NorviTech
/
arrmes-qwen3-0.6b

MLX
Safetensors
qwen3
lora
tool-calling
arrmes
4-bit precision
Model card Files
xet
 Community

Arrmes Qwen3-0.6B v9 Mixed-Precision QLoRA Adapter (v1.6.0)

Unassisted FP16 LoRA adapter trained on a clean 4-bit Qwen3-0.6B base with <think> assistant-prefix masking and balanced hard-negative/clarification/refusal curriculum (TrainPool-v9).

Arrmes v1.6.0 Accuracy Comparison

Artifact & Unassisted Benchmark Summary (LocalModel, backend="mlx-lm")

  • v9/adapters.safetensors SHA-256: 257555fcb31fb9f7c1fcd48b4c52f5f14fa3d705c425cdf8b5502b6befca60b4
  • Zero Router-Overwrite Guarantee: LocalModel (mlx-lm) executes genuine unassisted MLX decoding in raw_unverified (Track A) and deterministic schema/grounding verification in raw (Track B).
  • Template-Disjoint acceptance-v7 v2 Positives (N=150): Track A Raw 138/150 (92.0%), Track B Verified 147/150 (98.0%, Wilson 95% LB 94.29%)
  • Template-Disjoint acceptance-v7 Hard Negatives (N=150): Track A Raw 67/150 (44.67%), Track B Verified 141/150 (94.0%, Wilson 95% LB 88.99%)
  • Frozen overnight-v1 (N=80): Track A Raw 64/80 (80.0%), Track B Verified 72/80 (90.0%)
  • Apple M4 Pro Metal Warm Latency: p50 = 206.65 ms, p95 = 231.93 ms (340.5 MiB combined 4-bit base + FP16 LoRA footprint).
Downloads last month
665
Safetensors
Model size
0.6B params
Tensor type
U32
·
BF16
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for NorviTech/arrmes-qwen3-0.6b

Base model

Qwen/Qwen3-0.6B-Base
Finetuned
Qwen/Qwen3-0.6B
Adapter
(737)
this model

Collection including NorviTech/arrmes-qwen3-0.6b