Qwen2.5-1.5B-ES-math

Qwen/Qwen2.5-1.5B (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 both the base model and an RL baseline) are in that repo's README.

Training

Param	Value
Base model	Qwen/Qwen2.5-1.5B
Method	Evolution Strategies (full-rank, es-at-scale)
Task	math
sigma	0.001
alpha (lr)	auto (sigma/2 = 0.0005)
Population size	32
Iterations	50
Train dataset	math_lvl3to5_8k (MATH levels 3-5, matches SimpleRL-Zoo's training set)
Batch size / mini-batch size	256 / 256
Max tokens	2048
Hardware	8x RTX 4090 48GB
Training wall-clock	3h 22m 35s

Usage

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

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zocrate/Qwen2.5-1.5B-ES-math")
tokenizer = AutoTokenizer.from_pretrained("zocrate/Qwen2.5-1.5B-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.