Upload recipes/qwen3.8-flash-next.yaml with huggingface_hub

recipes/qwen3.8-flash-next.yaml

@@ -0,0 +1,76 @@
+# Calibration recipe for Qwen/Qwen3.8-Flash-Next.
+# Generated by foundry make_recipe. Everything under model: was read from the
+# hub, not typed. The shares below are the one human judgement call: they say
+# what this model is for, which no config file knows.
+#
+# Architecture facts that matter here:
+#   hidden_size = 2560 (%256 = 0)
+#   moe_intermediate_size = 640 (%256 = 128)
+#
+# Shares:
+#   agentic 25      tool markup is a dense token set natural text never emits,
+#                   so without real weight the imatrix has no statistics for it
+#   code 18         the agentic evals in this class are code shaped
+#   reasoning 15    the thinking channel is on the default inference path here
+#   multilingual 14 non-Latin embedding rows are only exercised by real
+#                   multilingual text
+#   longctx 12      no sliding window, so long range behaviour is only exercised by long documents
+#   vocab_sweep 10  regenerated against THIS tokenizer, 248320 rows. A sweep
+#                   built for another model is meaningless here
+#   structured 5    JSON, YAML, TOML, SQL, as files and inside assistant turns
+#   graphics 3      kept small for continuity with the other builds
+
+name: qwen3.8-flash-next
+
+model:
+  repo: Qwen/Qwen3.8-Flash-Next
+  tokenizer: /src/tokenizer.json
+  gguf: /gguf/qwen3.8-flash-next-bf16.gguf
+  vocab_size: 248320
+  layers: 48
+  sliding_window: null
+  context_length: 262144
+  llama_cpp_commit: FILLED_AT_BUILD_TIME
+
+seed: 20260814
+
+chat:
+  # auto_fmt drives this model's own chat_template through transformers, so
+  # byte equality with inference holds by construction. Install it once with
+  # install_auto_fmt; no per-model renderer is needed.
+  format: auto
+  add_bos_per_document: false
+
+calib_train:
+  target_tokens: 5000000
+  max_doc_fraction:
+    default: 0.05
+    longctx: 0.12
+  shares:
+    agentic: 25
+    code: 18
+    reasoning: 15
+    multilingual: 14
+    longctx: 12
+    vocab_sweep: 10
+    structured: 5
+    graphics: 3
+
+calib_longctx:
+  target_tokens: 750000
+  doc_tokens_min: 16384
+  doc_tokens_max: 32768
+  sources: [longctx]
+
+exclude:
+  # Anything pre-rendered with another model's markup. Those byte sequences are
+  # not special tokens for this tokenizer, so they would calibrate on text this
+  # model never sees.
+  render: [dsv4, nemotron, muse-glimmer]
+  provenance_key: excluded_from_builds
+
+notes:
+  imatrix: >
+    llama-imatrix defaults to parse_special = false. Without --parse-special the
+    chat markup in calib_train.txt is tokenised as literal punctuation and the
+    agentic and reasoning slices calibrate on text the model never sees.