# Code-heavy variant of flint-moe for Flint (a coding-agent model): an experiment, not a house recipe.
# Same markup, vocabulary and build as flint-moe; only the shares move toward what Flint is used for
# (repository work, tool calls, tests). The question it answers: does an imatrix from the model's own use
# lower KLD on the code eval set at the low rungs without losing on neutral? Compared rung for rung against
# flint-moe on one stand (same reference logits, eval sets, llama.cpp build).
#
# Shares (sum 100): agentic and code take most; multilingual and vocab_sweep keep enough to reach rarely routed
# experts (256 experts, 6 active: at 5M tokens ~117k tokens per expert on average).
name: flint-code

model:
  repo: RecursiveSI/flint
  tokenizer: /src/tokenizer.json
  gguf: /gguf/flint-bf16.gguf
  vocab_size: 150144
  layers: 49
  sliding_window: 513
  context_length: 262144
  llama_cpp_commit: '19ab5a5e6573'

seed: 20261022

chat:
  format: auto
  add_bos_per_document: false

calib_train:
  target_tokens: 5000000
  max_doc_fraction:
    default: 0.05
    longctx: 0.12
  shares:
    agentic: 30
    code: 30
    reasoning: 10
    multilingual: 8
    longctx: 8
    vocab_sweep: 6
    structured: 7
    graphics: 1

calib_longctx:
  target_tokens: 750000
  doc_tokens_min: 16384
  doc_tokens_max: 32768
  sources: [longctx]

exclude:
  render: [dsv4, nemotron, muse-glimmer]
  provenance_key: excluded_from_builds

notes:
  imatrix: >
    Same parse_special note as flint-moe.