Downloads · 30 days
8
32% of all-time downloads
Qwen2.5-7B-ES-math-iter100
zocrate/Qwen2.5-7B-ES-math-iter100
Qwen2.5-7B-ES-math-iter100 is a machine learning model from zocrate. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
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).
All-time downloads
25
Public
Parameters
7.6B
15.2 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors15.2 GB · 100%
At a glance
- License
- apache-2.0
- Model type
- qwen2
- Access
- Public
- Created
- Aug 16, 2026
- Updated
- Aug 16, 2026
- SHA
- 110510f4
Base models
Model card
From the Hugging Face model README
Qwen2.5-7B-ES-math-iter100
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).
This is the 100-iteration continuation of
zocrate/Qwen2.5-7B-ES-math
(the Experiment 2 / 50-iteration checkpoint). Same σ = 0.001, same data, 50
further iterations. The extra training bought nothing: population-mean reward
went 0.664 → 0.645, every in-loop eval stayed inside noise, GPQA-diamond
dropped 3.0 points. Prefer the iter50
checkpoint unless you
specifically want this longer run.
Part of the ES-capacity project. This checkpoint was not put through the pass@k suite; numbers below are the trainer's in-loop eval and GPQA only. Full write-up is in that repo's README ("Additional runs" / "Run length").
Training
| Param | Value |
|---|---|
| Base model | Qwen/Qwen2.5-7B |
| Method | Evolution Strategies (full-rank, es-at-scale) |
| Task | math |
| sigma | 0.001 |
| alpha (lr) | auto (sigma/2 = 0.0005) |
| Population size | 32 |
| Iterations | 51–101 (resumed from iter50 with --checkpoint / --start-iteration 51) |
| 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 |
| Seed | 42 |
| Hardware | 8x RTX 4090 48GB |
| Continuation wall-clock | 4.20h (sum of per-iteration times) |
What changed vs iter50
| Measure | After 50 iters | After 100 iters | Δ |
|---|---|---|---|
| Training reward (population mean) | 0.664 | 0.645 | −0.019 |
| In-loop AIME | 6.67% | 6.67% | ±0.00 |
| In-loop AMC | 34.94% | 37.35% | +2.41 |
| In-loop MATH500 | 73.00% | 73.40% | +0.40 |
| In-loop Minerva | 37.87% | 37.87% | ±0.00 |
| In-loop OlympiadBench | 36.15% | 36.00% | −0.15 |
| GPQA-diamond | 31.8% | 28.8% | −3.03 |
The in-loop suite is a different harness from the project's pass@k numbers (single-sample pass@1, trainer prompt/sampling). Treat ~±3 points as noise on the mid-size benchmarks.
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-iter100")
tokenizer = AutoTokenizer.from_pretrained("zocrate/Qwen2.5-7B-ES-math-iter100")
Prompted and evaluated with the qwen-boxed template (see
limit-of-RLVR).
License
Apache 2.0, inherited from the base model.
- Type
- qwen2
- License
- apache-2.0
- Created
- Aug 16, 2026
- Updated
- Aug 16, 2026