ssurface/cot-dialect-qwen3-4b-thinking-sft-l2

Qwen3-4B-Thinking — L2 dialect (Structured shorthand)

A LoRA adapter that makes Qwen/Qwen3-4B-Thinking-2507 reason at compression level L2 — compressed prose / bulleted steps.

Results

GSM8K test (n=1317), greedy decoding, single-turn, no exemplars, no self-consistency.

Training data

GSM8K train, re-expressed at level L2 by a teacher model: 6950 examples, median chain length 140 characters inside <think>.

Across the family the median chain runs from 532 characters at L1 to 16 at L5 — a 33x span. An L2 chain looks like this:

- Sandoval: 12
- Hawkins: 12 / 2 = 6
- Sloan: 12 + 10 = 22
- Total: 12 + 6 + 22 = 40

Training setup

Loss is on the completion only, with prompt lengths precomputed at load time rather than found by pattern search — the pattern-search collator silently masked nothing, which let the base model's tool-calling prior leak into the chains.

Usage

Solve this using Level 2 (Concise).
Problem: {your problem}

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Thinking-2507", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "ssurface/cot-dialect-qwen3-4b-thinking-sft-l2")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Thinking-2507")

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

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