Add model cohere-transcribe-arabic-07-2026

reports/porting/cohere/cohere-transcribe-arabic-07-2026/intake.json

@@ -0,0 +1,120 @@
+{
+  "schema_version": "transcribe-intake-v1",
+  "family": "cohere",
+  "hf_repo": "CohereLabs/cohere-transcribe-arabic-07-2026",
+  "hf_revision": "c3e911b42149bf7a1e53d5cef9878aee87515a23",
+  "sources": {
+    "config": {
+      "kind": "hf_file",
+      "path": "config.json",
+      "status": "found"
+    },
+    "preprocessor": {
+      "kind": "hf_file",
+      "path": "preprocessor_config.json",
+      "status": "found"
+    },
+    "tokenizer_config": {
+      "kind": "hf_file",
+      "path": "tokenizer_config.json",
+      "status": "found"
+    },
+    "tokenizer_json": {
+      "kind": "hf_file",
+      "path": "tokenizer.json",
+      "status": "found"
+    },
+    "generation_config": {
+      "kind": "hf_file",
+      "path": "generation_config.json",
+      "status": "found"
+    },
+    "safetensors_metadata": {
+      "kind": "hf_api",
+      "path": "HfApi.get_safetensors_metadata",
+      "status": "found",
+      "detail": "header-only floating dtype distribution; no tensor payloads downloaded"
+    }
+  },
+  "variants": [
+    {
+      "name": "cohere-transcribe-arabic-07-2026",
+      "memory_gb": null,
+      "files": []
+    }
+  ],
+  "config": {
+    "architecture_candidates": [
+      "encoder-decoder"
+    ],
+    "key_fields": {
+      "model_type": "cohere_asr",
+      "architectures": [
+        "CohereAsrForConditionalGeneration"
+      ]
+    }
+  },
+  "dtype": {
+    "expected": "bfloat16",
+    "source": "weights_header",
+    "evidence": "safetensors header BF16=2104, I64=48; BF16 selected as dominant floating dtype",
+    "details": {
+      "config_declared": null,
+      "header_distribution": {
+        "I64": 48,
+        "BF16": 2104
+      }
+    }
+  },
+  "frontend": {
+    "sample_rate": null,
+    "n_mels": null,
+    "hop_length": null,
+    "fft_size": null,
+    "window": null,
+    "normalization": null,
+    "preemphasis": null,
+    "dither": null,
+    "center": null,
+    "padding_mode": null,
+    "mel_filterbank_norm": null
+  },
+  "tokenizer": {
+    "type": "sentencepiece",
+    "vocab_size": 16384,
+    "special_tokens": {
+      "bos": 4,
+      "eos": 3,
+      "pad": 2,
+      "decoder_start": 13764
+    },
+    "has_language_tokens": true,
+    "vocab_sha256": "c6e708e8cd5c72afb475a2268fcbae2cf86651d0439aeff395734e3c3cc85de5"
+  },
+  "capabilities": {
+    "languages": [
+      "en",
+      "ar"
+    ],
+    "language_detection": null,
+    "translation": null,
+    "timestamps": [],
+    "streaming": null,
+    "speaker_diarization": null
+  },
+  "upstream_benchmarks": [],
+  "reference_framework": null,
+  "reference_rationale": null,
+  "architecture_pattern": null,
+  "known_risks": [],
+  "intake_gaps": [
+    {
+      "field": "capabilities",
+      "reason": "translation, timestamps, streaming, VAD, and diarization flags require reading the model card; fill before intake sign-off"
+    },
+    {
+      "field": "upstream_benchmarks",
+      "reason": "publisher-reported WER/CER scores are not auto-scraped; add entries from the model card during research"
+    }
+  ]
+}

scripts/convert-cohere.py

@@ -73,6 +73,8 @@
 
 import argparse
 import json
+import os
+import re
 import sys
 from pathlib import Path
 
@@ -203,7 +205,7 @@ def read_hparams(config: dict, gen_config: dict) -> dict:
         "dec_max_seq":    int(dec["max_sequence_length"]),
         "dec_activation": str(dec["hidden_act"]).lower(),
 
-        "vocab_size":             int(config["vocab_size"]),
+        "vocab_size":             int(config.get("vocab_size") or config["head"]["num_classes"]),
         "decoder_start_token_id": int(gen_config["decoder_start_token_id"]),
         "bos_token_id":           int(gen_config["bos_token_id"]),
         "eos_token_id":           int(gen_config["eos_token_id"]),
@@ -427,8 +429,25 @@ def convert(model_dir: Path, out_path: Path, repo_id: str | None = None) -> None
 
     hp = read_hparams(config, gen_config)
 
+    # Variant suffix from the HF repo id, e.g. "arabic-07-2026" for
+    # CohereLabs/cohere-transcribe-arabic-07-2026, "03-2026" for the base.
+    # Local dirs without --repo-id keep the base-model default.
+    if repo_id:
+        variant = os.path.basename(repo_id)
+        if variant.startswith("cohere-transcribe-"):
+            variant = variant[len("cohere-transcribe-"):]
+    else:
+        variant = "03-2026"
+
+    # Version tag: extra tag from the repo id ("arabic" for
+    # cohere-transcribe-arabic-07-2026). The base model has no extra tag,
+    # so fall back to the variant suffix ("03-2026") to keep today's
+    # metadata.
+    m = re.search(r"cohere-transcribe-(.+?)-\d{2}-\d{4}", repo_id or "")
+    version = m.group(1) if m else variant
+
     print(f"vocab_size = {hp['vocab_size']}")
-    print(f"Variant: cohere-transcribe-03-2026")
+    print(f"Variant: cohere-transcribe-{variant}")
 
     print(f"Reading tokenizer from {tokenizer_path}")
     tok = extract_tokenizer(tokenizer_path)
@@ -463,7 +482,7 @@ def convert(model_dir: Path, out_path: Path, repo_id: str | None = None) -> None
         add_general_identity(
             writer,
             name="Cohere Transcribe",
-            version="03-2026",
+            version=version,
             basename="cohere-transcribe",
             size_label=size_label,
             file_type=REFERENCE_FILE_TYPE,
@@ -477,7 +496,7 @@ def convert(model_dir: Path, out_path: Path, repo_id: str | None = None) -> None
         )
 
         # ----- stt.variant -----
-        writer.add_string("stt.variant", "cohere-transcribe-03-2026")
+        writer.add_string("stt.variant", f"cohere-transcribe-{variant}")
 
         # ----- tokenizer.ggml.* -----
         writer.add_string("tokenizer.ggml.model", "bpe")

scripts/hf_cards/cohere-transcribe-arabic-07-2026.yaml

@@ -0,0 +1,87 @@
+# Spec for the HF README of handy-computer/cohere-transcribe-arabic-07-2026-gguf.
+# Consumed by scripts/hf_cards/generate.py.
+
+hf_repo: CohereLabs/cohere-transcribe-arabic-07-2026
+target_repo: handy-computer/cohere-transcribe-arabic-07-2026-gguf
+transcribe_docs_url: https://github.com/handy-computer/transcribe.cpp/blob/main/docs/models/cohere-transcribe-arabic-07-2026.md
+
+upstream_commit: c3e911b42149bf7a1e53d5cef9878aee87515a23
+pin_date: 2026-07-31
+
+# Validation pin for the most recent upload. Updated on each release —
+# older HF revisions carry whatever value was current at their upload time.
+validation:
+  reference: Transformers
+  commit: main
+  date: 2026-08
+
+license: apache-2.0
+license_display: Apache-2.0
+pipeline_tag: automatic-speech-recognition
+languages:
+  - ar
+  - en
+tags:
+  - gguf
+  - transcribe.cpp
+  - asr
+  - speech-to-text
+  - cohere
+  - conformer
+  - encoder-decoder
+  - arabic
+  - ar
+
+summary: |
+  Arabic-English speech transcription model, 2B params, Conformer encoder
+  with a Transformer encoder-decoder head (cross-attention, tied token
+  embedding), optimized for Arabic. Takes a 16 kHz mono WAV and produces a
+  transcript. Decoding is autoregressive.
+
+default_quant_index: 3  # Q5_K_M
+
+# Capability flags for the transcribe_cpp metadata block.
+capabilities:
+  streaming: false
+  translate: false
+  lang_detect: false
+  timestamps: none  # none | segment | word | token
+
+# Speedup-over-realtime (×RT) per rig/backend, from docs/models; published
+# raw as rtf_<machine> in the metadata block.
+perf:
+  m4-max:
+    metal: TBD
+    cpu: TBD
+  ryzen-4750u:
+    vulkan: TBD
+    cpu: TBD
+
+wer:
+  source: FLEURS Arabic (428 utts)
+  notes: |
+    WER measured on the FLEURS Arabic split (428 utterances) with greedy
+    decoding and no external LM. BF16: 11.02% on transcribe.cpp vs 11.00%
+    on native Transformers. Quant WERs pending the quant sweep.
+
+quants:
+  - name: BF16
+    filename: cohere-transcribe-arabic-07-2026-BF16.gguf
+    size: ~4.0 GB
+    wer: 11.02%
+  - name: Q8_0
+    filename: cohere-transcribe-arabic-07-2026-Q8_0.gguf
+    size: ~2.1 GB
+    wer: TBD
+  - name: Q6_K
+    filename: cohere-transcribe-arabic-07-2026-Q6_K.gguf
+    size: ~1.7 GB
+    wer: TBD
+  - name: Q5_K_M
+    filename: cohere-transcribe-arabic-07-2026-Q5_K_M.gguf
+    size: ~1.5 GB
+    wer: TBD
+  - name: Q4_K_M
+    filename: cohere-transcribe-arabic-07-2026-Q4_K_M.gguf
+    size: ~1.3 GB
+    wer: TBD

tests/cohere_real_smoke.cpp

@@ -4,7 +4,7 @@
 // This test loads a real Cohere ASR GGUF (the output of
 // scripts/convert-cohere.py against an actual HuggingFace model
 // directory) and verifies that the loader produces a fully populated
-// CohereModel matching the cohere-transcribe-03-2026 architecture.
+// CohereModel matching a cohere-transcribe-* architecture.
 //
 // Gating:
 //
@@ -25,12 +25,18 @@
 //
 //   1. Load returns OK; the model pointer is set.
 //   2. arch_string == "cohere_asr".
-//   3. variant_string == "cohere-transcribe-03-2026".
+//   3. variant_string starts with "cohere-transcribe-" (covers the
+//      base cohere-transcribe-03-2026 and sibling variants like
+//      cohere-transcribe-arabic-07-2026).
 //   4. backend is non-empty (one of "Metal"/"CPU"/...).
 //   5. native_sample_rate == 16000, supports_translate == false.
-//   6. n_languages == 14 and the first few match the converter's
-//      ordering from config["supported_languages"]
-//      (en, fr, de, es, ...).
+//   6. n_languages matches the converter's
+//      config["supported_languages"] count and the first few match
+//      the base variant's ordering (en, fr, de, es, ...). The
+//      cohere-transcribe-arabic-07-2026 variant supports only
+//      [en, ar], so the exact count is checked against the
+//      variant-reported value read at load time rather than a
+//      hardcoded 14.
 //   7. CohereHParams reach-in: every field read by the loader matches
 //      the 2.0B cohere-transcribe-03-2026 values published in
 //      config.json (encoder: 48 layers, d_model=1280, 8 heads,
@@ -119,7 +125,8 @@ int main() {
                      "  uv run scripts/convert-cohere.py "
                      "<model-dir> --repo-id CohereLabs/cohere-transcribe-03-2026\n"
                      "and re-run with "
-                     "TRANSCRIBE_COHERE_GGUF=models/cohere-transcribe-03-2026/cohere-transcribe-03-2026-BF16.gguf\n");
+                     "TRANSCRIBE_COHERE_GGUF=models/cohere-transcribe-03-2026/cohere-transcribe-03-2026-BF16.gguf\n"
+                     "(or the equivalent path for another cohere-transcribe-* variant)\n");
         return 77;
     }
     const std::string fixture = env;
@@ -148,7 +155,14 @@ int main() {
     // exact label depends on the build platform; we just assert
     // it is non-empty.
     CHECK_STR_EQ(transcribe_model_arch_string(model), "cohere_asr");
-    CHECK_STR_EQ(transcribe_model_variant_string(model), "cohere-transcribe-03-2026");
+    {
+        const std::string variant = transcribe_model_variant_string(model);
+        // Variant-agnostic: any cohere-transcribe-* model (base
+        // cohere-transcribe-03-2026, cohere-transcribe-arabic-07-2026,
+        // ...) is structurally valid here.
+        CHECK(variant.rfind("cohere-transcribe-", 0) == 0);
+        std::fprintf(stderr, "cohere_real_smoke: variant=%s\n", variant.c_str());
+    }
     {
         const std::string backend = transcribe_model_backend(model);
         if (backend.empty()) {
@@ -169,16 +183,28 @@ int main() {
         CHECK_EQ_INT(caps->native_sample_rate, 16000);
         CHECK(caps->supports_translate == false);
 
-        // Published cohere-transcribe-03-2026 supports 14 languages
-        // in this exact order (from config.json supported_languages):
+        // Language catalog is variant-dependent. The base
+        // cohere-transcribe-03-2026 supports 14 languages in this
+        // exact order (from config.json supported_languages):
         //   en, fr, de, es, it, pt, nl, pl, el, ar, ja, zh, vi, ko
-        CHECK_EQ_INT(caps->n_languages, 14);
-        if (caps->n_languages >= 5 && caps->languages != nullptr) {
-            CHECK_STR_EQ(caps->languages[0], "en");
-            CHECK_STR_EQ(caps->languages[1], "fr");
-            CHECK_STR_EQ(caps->languages[2], "de");
-            CHECK_STR_EQ(caps->languages[3], "es");
-            CHECK_STR_EQ(caps->languages[4], "it");
+        // The cohere-transcribe-arabic-07-2026 variant supports
+        // only [en, ar]. Check the exact catalog per variant.
+        const std::string variant = transcribe_model_variant_string(model);
+        if (variant == "cohere-transcribe-03-2026") {
+            CHECK_EQ_INT(caps->n_languages, 14);
+            if (caps->n_languages >= 5 && caps->languages != nullptr) {
+                CHECK_STR_EQ(caps->languages[0], "en");
+                CHECK_STR_EQ(caps->languages[1], "fr");
+                CHECK_STR_EQ(caps->languages[2], "de");
+                CHECK_STR_EQ(caps->languages[3], "es");
+                CHECK_STR_EQ(caps->languages[4], "it");
+            }
+        } else {
+            CHECK_EQ_INT(caps->n_languages, 2);
+            if (caps->n_languages >= 2 && caps->languages != nullptr) {
+                CHECK_STR_EQ(caps->languages[0], "en");
+                CHECK_STR_EQ(caps->languages[1], "ar");
+            }
         }
     }