{ "schema_version": "transcribe-intake-v1", "family": "gigaam", "hf_repo": "ai-sage/GigaAM-v3", "hf_revision": "c7f128b8accdd9624df905e5c2d7b7a48c27c0d8", "sources": { "config": { "kind": "hf_file", "path": "config.json", "status": "found", "detail": "Hydra-style config under cfg.model.cfg: same encoder + frontend as the other gigaam-v3 branches; head=RNNTHead with decoder.num_classes=34 and joint.num_classes=34 (33 chars + 1 blank). decoding=RNNTGreedyDecoding with INLINE character vocabulary (no tokenizer.model file). model_name=v3_rnnt; model_class=rnnt." }, "preprocessor": { "kind": "hf_file", "path": "preprocessor_config.json|feature_extractor_config.json", "status": "missing", "detail": "Frontend declared inside config.json under cfg.model.cfg.preprocessor (n_mels=64, n_fft=320, hop=160, win=320, mel_scale=htk, mel_norm=null, center=false, SpecScaler log). Identical to all sibling branches." }, "tokenizer_config": { "kind": "hf_file", "path": "tokenizer_config.json", "status": "missing", "detail": "Charwise tokenizer; vocabulary is inline in config.json under cfg.model.cfg.decoding.vocabulary (33 strings: space + 32 lowercase Cyrillic letters а-я)." }, "tokenizer_model": { "kind": "hf_file", "path": "tokenizer.model", "status": "missing", "detail": "Branch has NO tokenizer.model file — `Tokenizer.charwise=True` path. Vocab is the inline 33-entry list." }, "tokenizer_json": { "kind": "hf_file", "path": "tokenizer.json", "status": "missing", "detail": "Not used." }, "generation_config": { "kind": "hf_file", "path": "generation_config.json", "status": "missing", "detail": "Decoding is RNN-T greedy; no HF generation_config." }, "model_weights": { "kind": "hf_file", "path": "pytorch_model.bin", "status": "found", "detail": "PyTorch pickle, 446,387,239 bytes (~446 MB). Smaller than e2e_rnnt's 449 MB because joint.num_classes=34 vs e2e_rnnt's 1025 (joint output projection is 320 -> 34 vs 320 -> 1025; predictor embedding is 34*320 vs 1025*320)." }, "safetensors_metadata": { "kind": "hf_api", "path": "HfApi.get_safetensors_metadata", "status": "missing", "detail": "Repo ships PyTorch pickle, not safetensors." }, "model_card": { "kind": "hf_api", "path": "https://huggingface.co/ai-sage/GigaAM-v3", "status": "found", "detail": "Single model card covers all 5 branches; v3_rnnt = v3_e2e_rnnt after upstream's post-processing strips punctuation+casing — both score the same in evaluation.md." }, "reference_modeling_code": { "kind": "reference_code", "path": "https://github.com/salute-developers/GigaAM (gigaam package: gigaam.load_model('v3_rnnt'))", "status": "found", "detail": "Author-first PyTorch implementation. License MIT. Same ConformerEncoder/RNNTHead/FeatureExtractor classes as e2e_rnnt; difference is num_classes (34 vs 1025) and Tokenizer.charwise path." } }, "variants": [ { "name": "gigaam-v3-rnnt", "memory_gb": 0.45, "files": [ "config.json", "modeling_gigaam.py", "pytorch_model.bin" ] } ], "config": { "architecture_candidates": [ "encoder-transducer" ], "key_fields": { "model_type": "gigaam", "model_class": "rnnt", "model_name": "v3_rnnt", "encoder.type": "ConformerEncoder", "encoder.feat_in": 64, "encoder.n_layers": 16, "encoder.d_model": 768, "encoder.n_heads": 16, "encoder.ff_expansion_factor": 4, "encoder.self_attention_model": "rotary", "encoder.pos_emb_max_len": 5000, "encoder.subsampling": "conv1d", "encoder.subs_kernel_size": 5, "encoder.subsampling_factor": 4, "encoder.conv_kernel_size": 5, "encoder.conv_norm_type": "layer_norm", "head.type": "RNNTHead", "decoder.pred_hidden": 320, "decoder.pred_rnn_layers": 1, "decoder.num_classes": 34, "joint.enc_hidden": 768, "joint.pred_hidden": 320, "joint.joint_hidden": 320, "joint.num_classes": 34, "decoding": "RNNTGreedyDecoding", "tokenizer.kind": "charwise", "tokenizer.vocab_size_plus_blank": 34, "tokenizer.vocab_inline": "33 entries: space + 32 lowercase Cyrillic letters (а б в г д е ж з и й к л м н о п р с т у ф х ц ч ш щ ъ ы ь э ю я; no ё)" }, "varying_across_variants": [ "model_class", "model_name", "head.type", "head.num_classes", "decoder.num_classes", "joint.num_classes", "tokenizer.kind", "tokenizer.vocab_inline" ] }, "dtype": { "expected": "float32", "source": "manual", "evidence": "pytorch_model.bin is 446 MB ≈ 111M params at F32. salute-developers/GigaAM PyTorch checkpoints are stored at F32.", "details": { "config_declared": null, "header_distribution": {} }, "expected_f32_tensors": [] }, "frontend": { "sample_rate": 16000, "n_mels": 64, "hop_length": 160, "fft_size": 320, "window": "hann_periodic", "normalization": "none", "preemphasis": null, "dither": null, "center": false, "padding_mode": null, "mel_filterbank_norm": "htk" }, "tokenizer": { "type": "other", "vocab_size": 33, "special_tokens": { "blank": 33 }, "has_language_tokens": false, "vocab_sha256": null }, "capabilities": { "languages": ["ru"], "language_detection": false, "translation": false, "timestamps": [], "streaming": false, "speaker_diarization": false }, "upstream_benchmarks": [ { "dataset": "Open Datasets (aggregate)", "language": "ru", "metric": "wer", "score": 2.6, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_RNNT row from the model card 5-domain WER summary. Same architecture and weights structure as v3_rnnt branch (lowercased Russian, no punctuation)." }, { "dataset": "Golos Farfield", "language": "ru", "metric": "wer", "score": 3.9, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_RNNT row from the model card." }, { "dataset": "Natural Speech", "language": "ru", "metric": "wer", "score": 6.9, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_RNNT row from the model card." }, { "dataset": "Callcenter", "language": "ru", "metric": "wer", "score": 9.5, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_RNNT row from the model card." }, { "dataset": "Aggregate (11-set average)", "language": "ru", "metric": "wer", "score": 8.3, "score_unit": "percent", "source": "https://github.com/salute-developers/GigaAM/blob/main/evaluation.md", "notes": "V3_RNNT macro-average WER over all 11 evaluation sets per evaluation.md." } ], "reference_framework": "author_repo_gigaam", "reference_rationale": "Same as the rest of the gigaam family: the `gigaam` PyPI package (MIT) is the canonical inference path. Load with `gigaam.load_model('v3_rnnt')`.", "architecture_pattern": "encoder-transducer", "known_risks": [ "Encoder + frontend are expected IDENTICAL to gigaam-v3-e2e-rnnt. Stage 3 must confirm encoder-tensor reuse. Head differs only in num_classes (34 vs 1025).", "Charwise tokenizer path: vocab is the inline 33-entry list inside config.json (`cfg.model.cfg.decoding.vocabulary`). NO tokenizer.model file. Detokenization is plain string join — `''.join(vocab[t] for t in tokens)`. Blank id = 33 = len(vocab). The convert script must read the inline vocab and emit it into the GGUF; the C++ loader must handle the charwise path distinctly from the SP path.", "Vocabulary covers space + 32 lowercase Cyrillic letters (а б в г д е ж з и й к л м н о п р с т у ф х ц ч ш щ ъ ы ь э ю я). Note: NO `ё` — typical of Russian ASR vocabs that fold `ё` into `е`. NO English letters, NO digits, NO punctuation, NO casing. Model output is lowercased Russian only.", "RNNTGreedyDecoding inner loop emits at most decoder_max_symbols steps per encoder frame; LSTM hidden state carries across frames. Same state-management risk as e2e_rnnt.", "Rotary positional embeddings inside Conformer attention — same element-ordering risk as the other variants.", "Conv1d subsampling, SpecScaler log scaling, center=false, mel_scale=htk, mel_norm=null — same gotchas as the sibling variants.", "Stage 7 FLEURS ru gate: FLEURS references include casing + punctuation; v3_rnnt outputs lowercased no-punct Russian. WER scorer must strip punctuation, lowercase, and fold `ё→е` on the reference side to match the model's vocab before computing WER. Without that normalization, WER will be artificially inflated.", "Long-form (>25s) path: same PyAnnote dependency — out of scope for the C++ port at parity." ], "intake_gaps": [ { "field": "tokenizer.vocab_sha256", "reason": "Computed at Stage 2 over the inline 33-entry charwise vocab. Detects vocab drift across the rnnt/ctc branches (these two are expected to share the same vocab list)." }, { "field": "dtype.expected_f32_tensors", "reason": "Filled at Stage 4 once we observe which norm/bias tensors must stay F32 for numerical match under non-F32 quant." }, { "field": "upstream_benchmarks.fleurs_ru", "reason": "FLEURS ru WER is not published by upstream. Stage 7 will measure it ourselves by running `gigaam.load_model('v3_rnnt').transcribe()` over `fleurs:ru` (with the lowercase + strip-punct + ё-fold reference normalization described in known_risks) and use the measured number as the upstream anchor." } ] }