GLM-4-32B-0414-Z1-SLERP

GLM-4-32B-0414-Z1-SLERP

A SLERP merge of GLM-4-32B-0414 (instruct) and GLM-Z1-32B-0414 (reasoning) at t=0.3 (~70% instruct / 30% reasoning). No Korean fine-tuning — this is an experimental base artifact, shared transparently.

TL;DR (honest)

On GSM8K (5-shot, 300 samples) this merge scores 94.0% vs 88.3% for its instruct parent — a +5.7pp gain (~2.4 sigma, statistically significant). It absorbs some of the reasoning parent's math skill while keeping the instruct parent's concise answering. This is the only benchmark run. Not a SOTA claim.

Results

*The reasoning parent's low score is an evaluation artifact: its long chain-of-thought is truncated at the default generation length — not its true ability.

Recipe

merge_method: slerp
base_model: zai-org/GLM-4-32B-0414
models:
  - model: zai-org/GLM-Z1-32B-0414
parameters:
  t: 0.3
dtype: bfloat16

Made with mergekit. Note: task-vector methods (dare_ties / ties with a base) collapsed this instruct+reasoning pair into incoherent multilingual/code output; only SLERP (direct interpolation, no base subtraction) yielded a coherent model.

Limitations

• Only GSM8K evaluated — no general, safety, or Korean benchmarks.
• Instruct parent diluted 30%; other capabilities may regress.
• No Korean fine-tuning (future work).
• Transparent reproducible reference, not a production/SOTA model.

Attribution

GLM-4-32B-0414 / GLM-Z1-32B-0414 (c) Zhipu AI (Z.ai), MIT License. Merged with mergekit (Arcee AI).