scrubster/dr-stein-colab-qwen3b-math-5k

PEFT/LoRA adapter for Qwen/Qwen2.5-3B-Instruct, fine-tuned on the slm-learning GAD-tool translation track.

• Trained on: 153 hand-curated (instruction, gad CLI command) pairs
• Adapter kind: lora (r=16, alpha=32)
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Run name: colab_qwen3b_math_5k
• Compute: colab-a100

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B-Instruct")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct")
model = PeftModel.from_pretrained(base, "scrubster/dr-stein-colab-qwen3b-math-5k")

Notes

The moonshot for this Colab session: 3B base + 5000 OpenMathInstruct-2 pairs. If math gains from base-size, this is where we'd see it. bs=4 grad_accum=4 (effective batch=16) fits comfortably on A100 40GB. Wall time estimate: 12 min train + ~5 min eval. Inference path for 3B locally: needs int8 (3GB) or stays remote. Acceptable for the math/reasoning track since each invocation is rare (compared to CLI which is per-command).