ssurface/cot-dialect-qwen3-4b-instruct-grpo-stop-l4

Qwen3-4B-Instruct — L4 dialect (Ultra-compact) · reward ablation stop

A LoRA adapter that makes Qwen/Qwen3-4B-Instruct-2507 reason at compression level L4 — semicolon-chained assignments.

This is an ablation, not one of the headline models: it is the same level trained under a different reward, published so the reward-design comparison in the paper can be rerun rather than taken on faith. For the main model at this level see ssurface/cot-dialect-qwen3-4b-instruct-grpo-l4.

Results

This adapter was not separately benchmarked. It exists as a training artefact for the ablation grid; the levels that carry reported numbers are the ones in the collection's core set.

Training data

GSM8K train, re-expressed at level L4 by a teacher model: 6976 examples, median chain length 41 characters inside <think>.

Across the family the median chain runs from 532 characters at L1 to 16 at L5 — a 33x span. An L4 chain looks like this:

K=18*2.5;D=8*4;T=K+D->T=77

Training setup

GRPO on top of the merged level-4 SFT model.

Engine	trl.GRPOTrainer on stock transformers, attention sdpa
Reward	correctness, format, sft_length, stop, gdpo
Loss type	dapo
Generations per prompt	8
Batch	64 x 1 accum
Max completion	256 tokens
Learning rate	1e-05
KL coefficient (beta)	0
Prompt set	/home/jovyan/.mlspace/envs/student_env/grpo/gsm8k_grpo_balanced_1k copy.json
Trained on	./merged_new_fixed/l4
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>
• sft_length — penalty measured against that row's own SFT chain length
• stop — penalizes failure to emit a stop token
• gdpo — normalizes each reward independently within the group before summing, so one component cannot swamp another

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 4 (Shorthand).
Problem: {your problem}

Stacks on the SFT model, not the raw base. Trained against the merged SFT model, so loading it straight onto Qwen/Qwen3-4B-Instruct-2507 will not reproduce the number above.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "ssurface/cot-dialect-qwen3-4b-instruct-sft-l4")   # 1. SFT for this level
model = model.merge_and_unload()
model = PeftModel.from_pretrained(model, "ssurface/cot-dialect-qwen3-4b-instruct-grpo-stop-l4")   # 2. this adapter
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-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).
• Ablation artefact. It was trained to answer one question about reward design and may be worse than the core model at the same level.

Citation

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