{ "schema_version": "transcribe-intake-v1", "family": "cohere", "hf_repo": "CohereLabs/cohere-transcribe-03-2026", "hf_revision": "76b8b23e8607f35f0265a23d481b338fb0e26aea", "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": "BF16=2104 (all floating tensors), I64=48 (integer buffers)" }, "model_card": { "kind": "hf_api", "path": "https://huggingface.co/CohereLabs/cohere-transcribe-03-2026", "status": "found", "detail": "language list, capability flags, and the discussion #28 note that drove the trust_remote_code=False decision" }, "reference_modeling_code": { "kind": "reference_code", "path": "models/cohere-transcribe-03-2026/{modeling_cohere_asr.py,processing_cohere_asr.py,tokenization_cohere_asr.py}, transcribe.cpp/scripts/convert-cohere.py, transcribe.cpp/src/arch/cohere/", "status": "found", "detail": "Local HF checkout includes the trust_remote_code modeling files; we do NOT use the remote code path for goldens (per family note + HF discussion #28), but it is the source of truth for shapes and frontend constants. Native Transformers (CohereAsrForConditionalGeneration in mainline transformers) is what the reference dumper actually runs." } }, "variants": [ { "name": "cohere-transcribe-03-2026", "memory_gb": 5.0, "files": ["model.safetensors", "config.json", "preprocessor_config.json", "tokenizer.json", "tokenizer.model"] } ], "config": { "architecture_candidates": ["encoder-decoder"], "key_fields": { "architectures": ["CohereAsrForConditionalGeneration"], "model_type": "cohere_asr", "vocab_size": 16384, "is_encoder_decoder": true, "max_seq_len": 1024, "max_audio_clip_s": 35, "encoder.type": "ConformerEncoder (FastConformer, NeMo-derived)", "encoder.n_layers": 48, "encoder.d_model": 1280, "encoder.n_heads": 8, "encoder.ff_expansion_factor": 4, "encoder.conv_kernel_size": 9, "encoder.subsampling": "dw_striding", "encoder.subsampling_factor": 8, "encoder.subsampling_conv_channels": 256, "encoder.self_attention_model": "rel_pos", "encoder.feat_in": 128, "encoder.untie_biases": true, "transf_decoder.hidden_size": 1024, "transf_decoder.inner_size": 4096, "transf_decoder.num_attention_heads": 8, "transf_decoder.num_layers": 8, "transf_decoder.hidden_act": "relu", "transf_decoder.pre_ln": true, "transf_decoder.max_sequence_length": 1024, "transf_decoder.learn_positional_encodings": false, "head.hidden_size": 1024, "head.num_classes": 16384, "head.activation": "relu", "decoding.strategy": "beam", "decoding.beam_size": 1, "prompt_format": "cohere_asr" }, "varying_across_variants": [] }, "dtype": { "expected": "bfloat16", "source": "weights_header", "evidence": "safetensors header BF16=2104 (all floating tensors), I64=48 (integer buffers); BF16 selected as dominant floating dtype. Cross-checked: converted GGUF is cohere-transcribe-03-2026-BF16.gguf and validate.py compare passes against native Transformers BF16 reference dumps.", "details": { "config_declared": null, "header_distribution": { "I64": 48, "BF16": 2104 } }, "expected_f32_tensors": [] }, "frontend": { "sample_rate": 16000, "n_mels": 128, "hop_length": 160, "fft_size": 512, "window": "hann_periodic", "normalization": "per_feature", "preemphasis": 0.97, "dither": 1e-05, "center": true, "padding_mode": "constant", "mel_filterbank_norm": "slaney" }, "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", "fr", "de", "es", "it", "pt", "nl", "pl", "el", "ar", "ja", "zh", "vi", "ko" ], "language_detection": false, "translation": false, "timestamps": ["none"], "streaming": false, "speaker_diarization": false }, "upstream_benchmarks": [], "reference_framework": "transformers", "reference_rationale": "Native Hugging Face Transformers (`CohereAsrForConditionalGeneration` in mainline transformers, `trust_remote_code=False`) is the canonical reference. The HF repo also ships a `trust_remote_code=True` modeling path, but Cohere has explicitly recommended the native path going forward and HF discussion #28 documents garbage generations on the remote-code path (https://huggingface.co/CohereLabs/cohere-transcribe-03-2026/discussions/28). Our reference dumper enforces `trust_remote_code=False`. NeMo is not a candidate: although the encoder is FastConformer-derived, the model is not packaged as a NeMo checkpoint.", "architecture_pattern": "encoder-decoder", "known_risks": [ "FastConformer encoder with depth-wise striding subsampling factor 8 — same family as Parakeet's encoder. preemph=0.97 applied before STFT; per-feature normalization; mismatches here silently degrade WER without changing tensor shapes.", "Encoder is large (48 layers, d_model=1280) — about 4x the Parakeet encoder by parameter count. Memory budgeting and Metal/Vulkan parity work need to account for this, not assume Parakeet sizing.", "Transformer decoder uses sinusoidal (`learn_positional_encodings=false`) absolute position encodings, pre-LN with a final extra LN (`pre_ln_final_layer_norm=true`). Off-by-one or post-LN variants are a common bring-up bug.", "Decoder is conditioned on a 10-token English-punctuation prompt: `[13764, 7, 4, 16, 62, 62, 5, 9, 11, 13]`. The `<|notimestamp|>` slot is hard-wired in `prompt_defaults`, which is why the family advertises max_timestamp_kind=NONE — there is no timestamp signal to recover.", "Multilingual prompt slots (`source_lang`, `target_lang`, `pnc`, `itn`, `diarize`, `emotion`, `timestamp`) live in the prompt template, not in separate task heads. Changing language requires substituting prompt slot tokens; the converter must round-trip the full prompt-format spec into GGUF so the C++ loader can rebuild it.", "Tokenizer is SentencePiece with custom Cohere-ASR added tokens (special slot tokens such as `<|en|>`, `<|notimestamp|>`, `<|pnc|>`). vocab_size=16384 includes these. Tokenizer round-trip must preserve the slot-token IDs exactly.", "Beam search size 1 (= greedy with a length-penalty hook); the converter records this but the C++ runner must not assume true beam>1 capacity is exercised.", "Cohere's published `trust_remote_code=True` path is known-broken (discussion #28). Do not import or compare against the remote-code generation path; only use it to read shapes." ], "intake_gaps": [ { "field": "upstream_benchmarks", "reason": "Cohere's model card does not publish a numerical WER/CER table at the time of intake; record entries from any subsequent technical report or blog post when available." }, { "field": "variants[].memory_gb", "reason": "Estimated from BF16 GGUF size; not measured against actual peak inference memory." } ] }