strands-qwen3-0.6b — Strands Agents expert (LoRA on Qwen3-0.6B)

LoRA adapter post-tuning Qwen/Qwen3-0.6B on the Strands Agents / Strands Robots corpus (same recipe as cagataydev/strands-qwen3-vl-2b and cagataydev/strands-gemma4-e2b). This is the smallest slow-weights backbone in the SLM family — designed to compose with the fast plastic layer from github.com/cagataycali/slm for test-time learning.

Results

metric	before	after
holdout NLL (strands corpus)	2.172	1.581 (Δ −0.59, −27%)

Training: 600 steps, LoRA r=32 (α=64, dropout 0.05) on q/k/v/o + gate/up/down, packed 1024-token blocks, bs 4 × accum 2, cosine LR 1e-4, bf16, single L40S.

Use with SLM (self-learning provider for Strands Agents)

from slm import SLM
from strands import Agent

model = SLM("cagataydev/strands-qwen3-0.6b")   # base auto-resolved + adapter merged
agent = Agent(model=model)
agent("How do I create a custom tool in Strands?")   # learns from every turn
model.save("experience.pt")                          # persist what it learned

Or plain transformers/peft:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B", dtype="bfloat16")
model = PeftModel.from_pretrained(base, "cagataydev/strands-qwen3-0.6b").merge_and_unload()
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")

SLM model family

repo	base	params	holdout NLL Δ
cagataydev/strands-qwen3-vl-2b	Qwen3-VL-2B-Instruct	2B	1.85 → ~1.0
cagataydev/strands-gemma4-e2b	Gemma 4 E2B (QAT mobile)	2B eff.	2.69 → 1.26
cagataydev/strands-qwen3-0.6b	Qwen3-0.6B	0.6B	2.17 → 1.58