Base model: mistralai/Mistral-7B-Instruct-v0.3
Task vectors:
- safety = wvnvwn/Mistral-7B-Instruct-v0.3-hhrlhf-v1, alpha = 0.7
- downstream = wvnvwn/Mistral-7B-Instruct-v0.3-gsm8k-v1, alpha = 0.30000000000000004

Hyperparameters

Method: dare
Scaling coefficient (lambda): 1.0
DARE drop rate: 0.9
dtype: bfloat16

Merge formula

$$\theta_\text{merged} = \theta_\text{base} + \lambda \cdot M\bigl(\alpha \cdot \tau_\text{safety} + (1-\alpha) \cdot \tau_\text{downstream}\bigr),$$

where tau_i = theta_i_sft - theta_base, and M(.) is the identity for TA, the Trim-Elect-Disjoint operator for TIES, and DARE's drop-and-rescale preprocessing (composed with TA) for DARE.

Reproduction

Produced by the GTM merge_cli (see commit history of the GTM repo). Full metadata is shipped as merge_metadata.json alongside the weights.

Intended use

Research only. This checkpoint is part of a sweep designed to probe the safety / downstream-utility Pareto frontier; individual α values are NOT recommended for deployment without further evaluation.

wvnvwn/Mistral-7B-Instruct-v0.3-gsm8k-dare-p0p9-lam1p0-a0p7-v1

Mistral-7B-Instruct-v0.3-gsm8k-dare-p0p9-lam1p0-a0p7-v1

Construction

Hyperparameters

Merge formula

Reproduction

Intended use

wvnvwn/
Mistral-7B-Instruct-v0.3-gsm8k-dare-p0p9-lam1p0-a0p7-v1