# TIES merge configuration for 4 specialized Qwen fine-tunes
# Goal:
# - Preserve strong domain expertise from each model
# - Reduce destructive interference between skills
# - Keep general reasoning stability from the base model
#
# Recommended for:
# - Qwen2 / Qwen2.5 instruction-tuned variants
# - Same architecture + same parameter count
# - Same tokenizer + same base checkpoint lineage

models:
  # ---------------------------
  # Math specialist
  # ---------------------------
  - model: cs-552-2026-claude-bots/math_model
    parameters:
      # High density because math capabilities are usually sparse
      # and easily lost during merging
      density:
        - filter: self_attn
          value: 0.72
        - filter: mlp
          value: 0.82

      # Strong contribution in reasoning-heavy blocks
      weight:
        - filter: self_attn
          value: 1.25
        - filter: mlp
          value: 1.15
        - value: 1.10

  # ---------------------------
  # Knowledge / factual model
  # ---------------------------
  - model: cs-552-2026-claude-bots/general_knowledge_model
    parameters:
      # Moderate density:
      # factual tuning tends to be more distributed
      density: 0.58

      # Slightly lower than math to avoid overwriting reasoning
      weight:
        - filter: self_attn
          value: 1.00
        - filter: mlp
          value: 0.95
        - value: 0.95

  # ---------------------------
  # Multilingual specialist
  # ---------------------------
  - model: cs-552-2026-claude-bots/multilingual_model
    parameters:
      # Language capabilities are often spread broadly,
      # so use reasonably high density
      density:
        - filter: embed_tokens
          value: 0.90
        - filter: self_attn
          value: 0.68
        - filter: mlp
          value: 0.62

      # Stronger influence on embeddings and attention
      weight:
        - filter: embed_tokens
          value: 1.30
        - filter: self_attn
          value: 1.10
        - value: 1.00

  # ---------------------------
  # Safety / alignment model
  # ---------------------------
  - model: cs-552-2026-claude-bots/safety_model
    parameters:
      # Lower density prevents excessive refusals
      # while still preserving alignment behavior
      density: 0.34

      # Important but intentionally constrained
      weight:
        - filter: self_attn
          value: 0.82
        - filter: mlp
          value: 0.72
        - value: 0.75

merge_method: ties

# IMPORTANT:
# Use the ORIGINAL shared pretrained base model
# from which all four fine-tunes were derived.
base_model: Qwen/Qwen3-1.7B

parameters:
  # Critical for TIES stability
  normalize: true

  # Helps reduce memory usage and improves masking behavior
  int8_mask: true

  # Trim very small parameter deltas
  # Good default for 4-way merges
  prune_threshold: 0.015

dtype: bfloat16

# Optional:
# tokenizer_source: base
# chat_template: auto