SOLID-Qwen3-4B-Instruct-2507
This is the step-125 checkpoint of SOLID (Solver-Informed Self-Distillation) built from Qwen/Qwen3-4B-Instruct-2507 for operations-research modeling and solver-backed answer generation.
The model was trained with GRPO and solver-informed token-level KL supervision. It uses the Gurobi-style StepORLM response template.
Evaluation
Each problem was sampled 64 times. maj@64 is majority-vote accuracy; pass@k uses the unbiased pass-at-k estimator. Objective correctness tolerance is 0.001.
| Dataset | maj@64 | pass@1 | pass@2 | pass@4 |
|---|---|---|---|---|
| OptMATH | 39.76 | 24.25 | 31.62 | 38.70 |
| MAMO-Complex | 33.00 | 29.51 | 35.71 | 41.34 |
| InOR | 53.00 | 43.92 | 50.85 | 56.64 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "JamesX421/SOLID-Qwen3-4B-Instruct-2507"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
The generated optimization code expects a compatible Gurobi environment for execution.