{
  "schema_version": "transcribe-intake-v1",
  "family": "granite_nar",
  "hf_repo": "ibm-granite/granite-speech-4.1-2b-nar",
  "hf_revision": "99a4df9007ac5682f9daa093fb7008ff606e9a5d",
  "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": "missing",
      "detail": "NAR decoding is single-pass; no autoregressive generation config needed"
    },
    "safetensors_metadata": {
      "kind": "hf_api",
      "path": "HfApi.get_safetensors_metadata",
      "status": "found",
      "detail": "header-only floating dtype distribution"
    },
    "reference_modeling_code": {
      "kind": "reference_code",
      "path": "ibm-granite/granite-speech-4.1-2b-nar repo files (configuration_nle.py, feature_extraction_nle.py, modeling_nle.py via trust_remote_code)",
      "status": "found",
      "detail": "Not in mainline transformers. auto_map points at in-repo Python modules. Loader must use trust_remote_code=True; the model card pins transformers==4.57.6 or 5.5.3 and flash-attn==2.8.3."
    },
    "model_card": {
      "kind": "hf_file",
      "path": "README.md",
      "status": "found"
    }
  },
  "variants": [
    {
      "name": "granite-speech-4.1-2b-nar",
      "memory_gb": null,
      "files": []
    }
  ],
  "config": {
    "architecture_candidates": [
      "encoder-ctc"
    ],
    "key_fields": {
      "architectures": [
        "NLENARDecoder"
      ],
      "model_type": "nle",
      "attn_implementation": "flash_attention_2",
      "scale_projected_embeddings": true,
      "encoder.num_layers": 16,
      "encoder.hidden_dim": 1024,
      "encoder.num_heads": 8,
      "encoder.dim_head": 128,
      "encoder.input_dim": 160,
      "encoder.output_dim": 348,
      "encoder.bpe_output_dim": 100353,
      "encoder.bpe_pooling_window": 4,
      "encoder.self_conditioning_layer": 8,
      "encoder.attn_type": "block",
      "encoder.conv_kernel_size": 15,
      "encoder.conv_expansion_factor": 2,
      "encoder.feedforward_mult": 4,
      "encoder.max_pos_emb": 512,
      "encoder.context_size": 200,
      "encoder.loss_lambda": 0.2,
      "projector.type": "nle_projector",
      "projector.block_size": 15,
      "projector.downsample_rate": 5,
      "projector.encoder_dim": 1024,
      "projector.hidden_size": 2048,
      "projector.llm_dim": 2048,
      "projector.mlp_ratio": 2,
      "projector.num_layers": 2,
      "projector.num_heads": 32,
      "projector.num_encoder_layers": 4,
      "projector.attn_bias": true,
      "projector.layernorm_eps": 1e-06,
      "projector.encoder_layer_indices": [4, 8, 12, -1],
      "text.model_type": "granite",
      "text.hidden_size": 2048,
      "text.num_hidden_layers": 40,
      "text.num_attention_heads": 16,
      "text.num_key_value_heads": 4,
      "text.vocab_size": 100352,
      "text.tie_word_embeddings": true,
      "text.rope_theta": 10000,
      "text.rms_norm_eps": 1e-05,
      "ctc_tokenizer.vocab_size": 348,
      "ctc_tokenizer.coverage": "ASCII 32-127 + Latin-1 supplement 128-255 + Katakana 0x30A1-0x30FB -> indices 256-347"
    },
    "varying_across_variants": [
      "model_type",
      "architectures",
      "projector.type",
      "scale_projected_embeddings",
      "audio_token_index",
      "text_config.tie_word_embeddings"
    ]
  },
  "dtype": {
    "expected": "bfloat16",
    "source": "config",
    "evidence": "config.dtype=bfloat16; safetensors header BF16=930 (all floating tensors) agrees.",
    "details": {
      "config_declared": "bfloat16",
      "header_distribution": {
        "BF16": 930,
        "I64": 16
      }
    },
    "expected_f32_tensors": []
  },
  "frontend": {
    "sample_rate": 16000,
    "n_mels": 80,
    "hop_length": 160,
    "fft_size": 512,
    "window": "hann_periodic",
    "normalization": "none",
    "preemphasis": null,
    "dither": 0.0,
    "center": true,
    "padding_mode": "reflect",
    "mel_filterbank_norm": "htk"
  },
  "tokenizer": {
    "type": "bpe",
    "vocab_size": 100352,
    "special_tokens": {
      "bos": 100257,
      "eos": 100257,
      "pad": 100256
    },
    "has_language_tokens": true,
    "vocab_sha256": "cd9faddbc5d020a41fb6dfa6924c090bf987ede9092bc088f6a7c66aef416af2"
  },
  "capabilities": {
    "languages": ["en", "fr", "de", "es", "pt"],
    "language_detection": false,
    "translation": false,
    "timestamps": ["none"],
    "streaming": false,
    "speaker_diarization": false
  },
  "upstream_benchmarks": [
    {
      "dataset": "LibriSpeech test-clean",
      "language": "en",
      "metric": "wer",
      "score": 1.29,
      "score_unit": "percent",
      "source": "model card",
      "notes": "Ref-dtype C++ port must score <= 1.30 to pass porting-7-wer."
    },
    {
      "dataset": "LibriSpeech test-other",
      "language": "en",
      "metric": "wer",
      "score": 2.75,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "Common Voice v15 en",
      "language": "en",
      "metric": "wer",
      "score": 6.50,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "Common Voice v15 de",
      "language": "de",
      "metric": "wer",
      "score": 4.73,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "Common Voice v15 es",
      "language": "es",
      "metric": "wer",
      "score": 4.02,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "Common Voice v15 fr",
      "language": "fr",
      "metric": "wer",
      "score": 7.17,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "Common Voice v15 pt",
      "language": "pt",
      "metric": "wer",
      "score": 2.57,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "MLS en",
      "language": "en",
      "metric": "wer",
      "score": 4.77,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "TED-LIUM v3",
      "language": "en",
      "metric": "wer",
      "score": 3.67,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "Earnings22",
      "language": "en",
      "metric": "wer",
      "score": 8.48,
      "score_unit": "percent",
      "source": "model card"
    },
    {
      "dataset": "GigaSpeech",
      "language": "en",
      "metric": "wer",
      "score": 10.12,
      "score_unit": "percent",
      "source": "model card"
    }
  ],
  "reference_framework": "transformers",
  "reference_rationale": "Reference is loaded via the transformers framework but using trust_remote_code: the auto_map in config.json points at in-repo configuration_nle.NLEConfig, feature_extraction_nle.NLEFeatureExtractor, and modeling_nle.NLENARDecoder. This is not in mainline transformers and has no separate author PyPI package (the code lives in the HF repo itself). The pinning is transformers==4.57.6 or 5.5.3 with flash-attn==2.8.3 per the model card. This variant is split into its own granite_nar family (separate from the granite audio-LLM family) because the non-autoregressive editor architecture (bidirectional LLM, no audio-token injection, single-pass CTC decode on LLM logits) does not share an implementation with GraniteSpeechForConditionalGeneration.",
  "architecture_pattern": "encoder-ctc",
  "known_risks": [
    "Non-autoregressive editor flow: Conformer encoder -> CTC head (vocab 348) emits a hypothesis with insertion slots -> NLE projector (MLP, not a Q-Former) lifts encoder features to 2048-d -> bidirectional Granite LLM (causal mask disabled) emits a single forward-pass logits tensor -> CTC greedy decode on the LLM output produces final text. There is no autoregressive token loop. architecture_pattern=encoder-ctc captures the encoder shape and final-decode op; the LLM-as-editor step is structurally novel and is not directly captured by any of the four canonical patterns.",
    "Encoder also emits a self-conditioned CTC at layer 8 (encoder.self_conditioning_layer=8) which feeds back into the residual stream. This is at training-time only; whether it is also applied at inference depends on modeling_nle.py and must be confirmed at oracle time.",
    "Projector consumes multi-layer encoder features (encoder_layer_indices=[4, 8, 12, -1]) -- four hidden states concatenated or attended jointly, not just the final layer. Stage 4 must wire up the multi-layer hidden-state collection.",
    "NLE projector is an MLP with attention (block_size=15, mlp_ratio=2, attn_bias=true, layernorm_eps=1e-06), not a Q-Former. Architecturally simpler than blip_2_qformer but with different layernorm eps (1e-06 vs the Q-Former's 1e-12).",
    "scale_projected_embeddings=true: the projector output is scaled before being fed to the LLM. The exact scale factor must be read from modeling_nle.py; a missing scale silently degrades accuracy.",
    "Bidirectional attention on a Granite LLM: causal mask is disabled. ggml graph for the LLM must not apply the standard causal triangular mask. Flash-attention reference requires flash-attn 2.8.3 with is_causal=False. Note that the LLM weights here are presumably LoRA-adapted from granite-4.0-1b-base for the editor task; whether LoRA deltas are merged into the safetensors or stored separately must be confirmed (the config does not set has_lora_adapter explicitly).",
    "Tokenizer vocab_size is 100352 here vs 100353 for the AR variants (the audio token at 100352 is the difference; NAR has no audio token because there is no LLM token-injection -- the LLM consumes projected encoder outputs as token embeddings directly).",
    "text_config.tie_word_embeddings=true here (vs false in the AR variants). Loader must wire up llama.cpp-style TENSOR_DUPLICATED fallback for lm_head.",
    "ctc_tokenizer is embedded inline in config (char2idx, 348 entries: ASCII 32-127 + Latin-1 supplement + Katakana). The CTC vocab covers Japanese kana but the supported_languages list omits Japanese -- the Katakana entries appear vestigial.",
    "feature_extractor is NLEFeatureExtractor (not GraniteSpeechFeatureExtractor) and uses flat preprocessor_config.json fields (no melspec_kwargs nesting). Numbers are identical to the AR variants' melspec.",
    "No autoregressive generation_config and no audio_token_index. The downstream runtime API must treat this variant differently -- a single forward pass producing a text string, not a streaming token loop."
  ],
  "intake_gaps": [
    {
      "field": "frontend.window/center/padding_mode/mel_filterbank_norm",
      "reason": "torchaudio MelSpectrogram defaults assumed (Hann periodic, center=True, reflect, htk, norm=None); preprocessor_config.json declares only the numeric fields. Oracle (porting-2) confirms by running the reference processor end-to-end."
    },
    {
      "field": "has_lora_adapter",
      "reason": "Not set in config; the model card hints at LoRA-adapted LLM weights for the editor task. Must be confirmed by inspecting safetensors tensor names at converter time."
    }
  ]
}