Datasets:
| # 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. | |