ssurface/cot-dialect-olmo3-7b-think-conditioned-grpo

Olmo-3-7B-Think — single conditioned model

The alternative design: one adapter asked to serve every verbosity level by naming the level in the prompt, instead of one adapter per level. This is the comparison that motivates the per-level family.

Results — GSM8K test (n=1317)

Requested level	Accuracy
L1	88.2%
L2	70.2%
L3	58.8%
L4	48.0%
L5	3.0%

The model holds up while the requested dialect stays close to what it settled on, then collapses at L5 (3.0%). A single set of weights cannot hold five dialects at once; asking for the extreme one gets neither the format nor the answer.

Training

Stage	GRPO on the merged conditioned SFT model
Conditioning	the level is named in the prompt, not selected by adapter
Engine	trl.GRPOTrainer on stock transformers, sdpa
LoRA	r=16, alpha=32
Hardware	1x NVIDIA A100 80GB

Usage

Solve this using Level {N} ({Verbose|Concise|Symbolic|Shorthand|Extreme}).
Problem: {your problem}

Stacks on the conditioned SFT model, not the raw base.

from transformers import AutoModelForCausalLM
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-olmo3-7b-think-conditioned-sft")
model = model.merge_and_unload()
model = PeftModel.from_pretrained(model, "ssurface/cot-dialect-olmo3-7b-think-conditioned-grpo")

Limitations

• A design comparison, not a recommended model. The per-level adapters in this collection are the ones the results are reported on.
• Single seed.