44 lines
1.4 KiB
YAML
44 lines
1.4 KiB
YAML
|
# Coral-v1.5-0.6B merge config (mergekit)
|
||
|
|
# Method: TIES — computes each finetune's task vector vs the base, trims to the
|
||
|
|
# strongest deltas, elects a consensus sign per parameter, then merges. Stable
|
||
|
|
# and quality-preserving (same method as CoralLM-1B). Stays at 0.6B params.
|
||
|
|
#
|
||
|
|
# Run:
|
||
|
|
# mergekit-yaml coral-v1.5-0.6b.yml ./coral-v1.5-0.6b-out --cuda
|
||
|
|
#
|
||
|
|
# All donors are Qwen3-0.6B finetunes -> shared tokenizer/arch, native <think>.
|
||
|
|
|
||
|
|
base_model: Qwen/Qwen3-0.6B # reference anchor, NOT a contributor
|
||
|
|
merge_method: ties
|
||
|
|
dtype: bfloat16
|
||
|
|
parameters:
|
||
|
|
normalize: true # keep merged magnitudes sane across 5 contributors
|
||
|
|
int8_mask: true # memory-light sign mask
|
||
|
|
|
||
|
|
models:
|
||
|
|
# --- Thinking / reasoning (the spicy stars, split across two sources) ---
|
||
|
|
- model: reaperdoesntknow/Qwen3-0.6B-Distilled-30B-A3B-Thinking-SFT
|
||
|
|
parameters:
|
||
|
|
weight: 0.30
|
||
|
|
density: 0.5
|
||
|
|
- model: MihaiPopa-1/Qwen-3-0.6B-Claude-4.7-Opus-Distilled
|
||
|
|
parameters:
|
||
|
|
weight: 0.30
|
||
|
|
density: 0.5
|
||
|
|
|
||
|
|
# --- Code ---
|
||
|
|
- model: suayptalha/Qwen3-0.6B-Code-Expert
|
||
|
|
parameters:
|
||
|
|
weight: 0.25
|
||
|
|
density: 0.5
|
||
|
|
|
||
|
|
# --- Uncensored / de-alignment (two sources, kept light so they don't dominate) ---
|
||
|
|
- model: DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored
|
||
|
|
parameters:
|
||
|
|
weight: 0.15
|
||
|
|
density: 0.5
|
||
|
|
- model: huihui-ai/Huihui-Qwen3-0.6B-abliterated-v2
|
||
|
|
parameters:
|
||
|
|
weight: 0.15
|
||
|
|
density: 0.5
|