Qwen2.5-7B-ES-math

Qwen2.5-7B-ES-math

Qwen/Qwen2.5-7B (base, not Instruct) fine-tuned with Evolution Strategies (gradient-free, population-based) on math reasoning, using the es-at-scale framework (arXiv:2509.24372).

Part of the ES-capacity project, which tests whether ES-based post-training preserves a base model's pass@k ceiling better than gradient-based RLVR (GRPO) — RLVR reliably raises pass@1 but tends to narrow the pass@k ceiling relative to the base model; full results (pass@k curves and solvable/unsolvable breakdowns across AIME24, MATH500, Minerva, and OlympiadBench, compared against the base model) are in that repo's README.

Same training recipe as the project's 1.5B run (zocrate/Qwen2.5-1.5B-ES-math), scaled to 7B with identical hyperparameters (kept unchanged deliberately, to isolate the effect of model scale).

Training

Usage

Standard transformers/vLLM Qwen2 checkpoint — no special loading code needed:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zocrate/Qwen2.5-7B-ES-math")
tokenizer = AutoTokenizer.from_pretrained("zocrate/Qwen2.5-7B-ES-math")

Prompted and evaluated with the qwen-boxed template (see limit-of-RLVR) during evaluation — same template family as SimpleRL-Zoo and the corresponding base model.

License

Apache 2.0, inherited from the base model.