初始化项目,由ModelHub XC社区提供模型
Model: NotHereNorThere/CoralLM-1b-raw Source: Original Platform
This commit is contained in:
42
mergekit_config.yml
Normal file
42
mergekit_config.yml
Normal file
@@ -0,0 +1,42 @@
|
||||
# CoralLM merge config (mergekit)
|
||||
# Method: TIES — trims each finetune's task vector to its strongest deltas,
|
||||
# elects a consensus sign per parameter, then merges. Stable and
|
||||
# quality-preserving on Llama-3.2-1B (same method bunnycore's popular
|
||||
# 1B/3B merges use). Swap merge_method to dare_ties to experiment.
|
||||
#
|
||||
# Run:
|
||||
# pip install mergekit
|
||||
# mergekit-yaml corallm_merge.yml ./corallm-merge --cuda
|
||||
#
|
||||
# Note: meta-llama/Llama-3.2-1B-Instruct is GATED. Before running:
|
||||
# huggingface-cli login (and accept the license on the model page)
|
||||
|
||||
base_model: meta-llama/Llama-3.2-1B-Instruct # reference anchor, not a contributor
|
||||
merge_method: ties
|
||||
dtype: float16
|
||||
parameters:
|
||||
normalize: true # keep merged magnitudes in check across 4 contributors
|
||||
int8_mask: true # memory-light sign mask
|
||||
|
||||
models:
|
||||
# --- Reasoning (split across two sources so it doesn't dominate) ---
|
||||
- model: EpistemeAI/Reasoning-Llama-3.2-1B-Instruct-v1.2
|
||||
parameters:
|
||||
weight: 0.30
|
||||
density: 0.5
|
||||
- model: Predacon/Pico-Lamma-3.2-1B-Reasoning-Instruct
|
||||
parameters:
|
||||
weight: 0.20
|
||||
density: 0.5
|
||||
|
||||
# --- Math / coding / logic ---
|
||||
- model: ai-nexuz/llama-3.2-1b-instruct-fine-tuned
|
||||
parameters:
|
||||
weight: 0.30
|
||||
density: 0.5
|
||||
|
||||
# --- General / creative / roleplay (loosens Meta's sterile alignment) ---
|
||||
- model: bunnycore/Llama-3.2-1B-General-Best
|
||||
parameters:
|
||||
weight: 0.25
|
||||
density: 0.5
|
||||
Reference in New Issue
Block a user