Scoring note. MATH answers are \boxed{}, and the harness that produced the first pass of these evals looked for GSM8K's #### n. That silently scored three of these models at ~0%% when they were near 60%%. Numbers here come from the project's LaTeX-aware grader, which normalizes equivalent forms (\frac{14}{3} == 14/3).

Training data

MATH training problems re-expressed at level L5 by a teacher model. MATH ships three levels rather than five — L1 anchor, L3 symbolic middle, L5 extreme — with the notation rules held identical to the GSM8K dialects and only the answer convention changed to \boxed{}.

Training setup

GRPO on top of the merged level-5 SFT model.


Engine	`trl.GRPOTrainer` on stock `transformers`, attention `sdpa`
Reward	`correctness`, `format`, `gr3`
Loss type	dapo
Generations per prompt	8
Batch	32 x 2 accum
Max completion	256 tokens
Learning rate	1e-05
KL coefficient (beta)	0.04
Prompt set	`math_grpo.json`
Trained on	`merged_math_olmo/l5`
LoRA	r=16, alpha=32
Hardware	1x NVIDIA A100 80GB

Reward components

correctness — +/- the gold solution's step count on an answer match, so harder problems are worth more
format — the response must be one <think>...</think> block then #### <answer>
gr3 — multiplicative length rescaling of the positive combined reward, floored at 0.3 -- it scales rewards that are already positive, so it cannot reorder correct above incorrect

Engine note. Stock transformers with sdpa attention, not a fused-kernel wrapper. The fused path produced adapters whose lora_B matrices were all zero — mathematically inert despite loading without error. Every adapter in this collection was verified lora_B != 0 before publishing; 13 that failed that check were withheld.

Usage

Solve this using Level 5 (Extreme).
Problem: {your problem}

Stacks on the SFT model, not the raw base. Trained against the merged SFT model, so loading it straight onto allenai/Olmo-3-7B-Think will not reproduce the number above.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("allenai/Olmo-3-7B-Think", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "ssurface/cot-dialect-math-olmo3-7b-think-sft-unfiltered-l5")   # 1. SFT for this level
model = model.merge_and_unload()
model = PeftModel.from_pretrained(model, "ssurface/cot-dialect-math-olmo3-7b-think-grpo-gr3-l5")   # 2. this adapter
tok = AutoTokenizer.from_pretrained("allenai/Olmo-3-7B-Think")

Limitations

Trained and evaluated on math word problems only.
Accuracy falls with problem difficulty, fastest at the compressed levels.
Single seed unless the repo name says otherwise; differences of a couple of points are within noise (95% half-width ~2.7 pp at n=1317, ~4.4 pp at n=500).

Citation

bibtex

@misc{cot-compression-dialects,
  title  = {Chain-of-Thought Compression Dialects},
  author = {Frolov, Anatolii},
  year   = {2026}
}