Training
- Base model: Qwen3.5-9B, thinking off.
- Method: DAPO in verl. The reward is 1 when the program the model writes runs and returns the certified
answer, otherwise 0.
- Data: problems generated from the templates of training fold 1 (29 of the 70 mechanisms, split by cluster),
each in its inside, outside and given versions.
- Settings: learning rate 1e-6, 128 problems × 8 samples per step, up to 120 steps; seed 42.
Use
The model was trained on one prompt format and should be given the same one. It answers with a Python program
that defines solve(). Run it with thinking disabled.
Load the model as you would its base model, Qwen/Qwen3.5-9B, and build the messages with
atmoscoder_rl.prompts.chat(problem) from the code repository (a fixed system prompt and a user template that
asks for solve()). Execution and grading of the returned program are in atmoscoder_rl/evaluation/.
Limitations
- This checkpoint is one fold and one seed. Numbers in the paper average both folds and two seeds, so this model
alone will not reproduce them exactly.
- Reinforcement learning alone rarely taught the model a formula its base model almost never produced. Where
training did not repair a mechanism, the model may still return the standard formula's value.
- It was trained and tested on problems from the study's templates and benchmark. It is not a general-purpose
assistant for atmospheric science.
Citation
@unpublished{ran2026regime, title = {Language models misapply atmospheric formulas until trained on a few worked examples}, author = {Ran, Maohao and Zhang, Yanting and Ma, Chendong and Song, Jun}, year = {2026}, note = {Manuscript under review}}
Licence
Apache 2.0, as for the base model Qwen3.5-9B.