Post-training Runs
We train frontier models on Surge datasets, then study what they learn: benchmark gains, transfer beyond the training distribution, and the behaviors that emerge.
We trained Nemotron 3.5 Super on four capabilities.
It improved across 10 benchmarks.
We trained Nemotron Super 3.5 on Surge data for agentic coding, long-horizon tool use, professional graphical reasoning, and instruction following.
Agentic Coding
pass@127.9% → 32.6%pass@186.9% → 92.1%pass@163.9% → 70.3%Long Horizon Tool Use
pass@111.7% → 23.2%Chart Understanding
With tools, pass@139.4% → 50.6%With tools, pass@179.35% → 85.80%Instruction Following
pass@127.3% → 35.5%pass@150.6% → 58.3%pass@172.3% → 76.8%pass@158.3% → 66.6%We trained a model on 2,400 coding tasks.
It improved across every SWE eval, again.
Training on 2,400 expert-built coding tasks improved Kimi K2.7 across every external SWE eval we measured. This builds on an earlier run with 1,700 tasks, with continued performance improvements.
PASS@131.0% → 48.7%pass@160.1% → 68.1%We trained a model on professional work.
It learned to produce deliverables it couldn't before.
Post-training on tasks from the DAYJOB companion set raised Kimi K2.7's mean rubric score on AA-Briefcase-Lite by +8.4 pp.
We trained a coding agent to think more like a staff engineer.
Post-training GLM-5.3 on 750 tasks from the sudo L7 companion set improved pass@1 on DeepSWE by +4.8 pp and Terminal-Bench 4 by +6.6 pp.
We trained a model on 1,700 coding tasks.
It got better at office work.
Training on Surge coding data improved GDPval-AA from 30.7% to 37.6%, even though GDPval-AA contains no coding tasks.
We trained a model on DAYJOB-style office work.
It got better at coding.
Training on the DAYJOB companion set, which contains no coding tasks, nonetheless improved Terminal-Bench 2.1 by +8.4pp.
zero coding tasks in training
GDPval success doubled after training on 1,400 GDP.pdf tasks
We post-trained Kimi K2.7 on 1,465 GDP.pdf companion tasks. The GDP.pdf holdout full-task pass rate rose +13.5pp and GDPval success more than doubled. The trained model used tools to find more of the relevant sources before creating deliverables, while reading less text overall.
held-out
held-out
We trained a model on 1,700 coding tasks.
It learned skills that generalized across every SWE eval.
Training on 1,700 Surge coding tasks improved Kimi K2.7 across every external SWE eval we measured. The biggest behavioral changes showed up in the complexity of real-world software engineering: carrying messy specifications through, testing beyond the obvious path, preserving existing behavior, and building its own checks when the environment did not provide one.
We trained a model on office work RL environments with no code. It got better at coding anyways.
We trained on long-horizon office tasks with no coding examples. SWE-Bench Pro still improved +5.7pp. Trajectory analysis pointed to a broader capability we call goal-directed execution: choosing useful subgoals, maintaining task state, staying faithful to the objective, and verifying completion.
zero coding tasks in training
We trained a model on single-turn ComplexConstraints. It also got better at multi-turn instruction following.
Every ComplexConstraints training example was single-turn, yet the largest gains appeared in multi-turn behaviors. Our hypothesis is that the underlying skill is shared: maintaining many interacting constraints without dropping them, whether they arrive in one dense prompt or accumulate over nine turns.
Qwen3-4B
Nemotron 3.5 Super
We trained a model on long-horizon tool-use environments. It learned agentic behaviors that transferred across every tool eval.
We trained Qwen3.5-122B-A10B on realistic, multi-tool agent environments. The gains transferred to Toolathlon, τ²-Bench, and BFCL-V4. Trajectory analysis showed better task closure and more efficient tool use across unfamiliar evaluation environments.
We trained a model on expert STEM problems. The gains transferred across every scientific reasoning benchmark.
We trained GLM-4.7 on expert-authored STEM problems. It nearly doubled performance on held-out Expert STEM tasks, with gains also appearing on HLE, FrontierScience, PubMedQA, and DeepDive. The result suggests that difficult expert problem solving can teach reasoning patterns that remain useful across different scientific domains.
held-out
text prompts
We trained a model on customer-support RL environments. It learned tool-use skills that generalized beyond.
We trained GLM-4.6 inside CoreCraft, a high-fidelity customer-support environment. The gains transferred beyond CoreCraft to BFCL, τ²-Bench Retail, and Toolathlon. The result suggests that realistic enterprise workflows can teach general agent skills that survive changes in tools, tasks, and evaluation environments.
held-out
See what high-quality datasets teach your models
All of these experiments use Surge datasets that are available off the shelf.
Train on the full dataset first. Pay only if it moves the metrics that matter.