{ "schema_version": "transcribe-intake-v1", "family": "gigaam", "hf_repo": "ai-sage/GigaAM-v3", "hf_revision": "15ef3b5a88da78f93134b3cb7f015c70aefa8946", "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=CTCHead with feat_in=768, num_classes=34 (33 chars + 1 blank). decoding=CTCGreedyDecoding with INLINE character vocabulary (no tokenizer.model file). model_name=v3_ctc; model_class=ctc." }, "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, no ё). IDENTICAL list to the rnnt branch — they share the lowercased-Russian-no-punct character vocab." }, "tokenizer_model": { "kind": "hf_file", "path": "tokenizer.model", "status": "missing", "detail": "Branch has NO tokenizer.model file — `Tokenizer.charwise=True` path." }, "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 CTC greedy; no HF generation_config." }, "model_weights": { "kind": "hf_file", "path": "pytorch_model.bin", "status": "found", "detail": "PyTorch pickle, 441,719,299 bytes (~441 MB). Smallest of the five branches because (a) CTC head is a single Conv1d projection vs RNN-T's predictor+joint, (b) num_classes=34 keeps the projection narrow." }, "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_ctc evaluated alongside v3_e2e_ctc in evaluation.md (both score the same after upstream post-processing)." }, "reference_modeling_code": { "kind": "reference_code", "path": "https://github.com/salute-developers/GigaAM (gigaam package: gigaam.load_model('v3_ctc'))", "status": "found", "detail": "Author-first PyTorch implementation. License MIT." } }, "variants": [ { "name": "gigaam-v3-ctc", "memory_gb": 0.44, "files": [ "config.json", "modeling_gigaam.py", "pytorch_model.bin" ] } ], "config": { "architecture_candidates": [ "encoder-ctc" ], "key_fields": { "model_type": "gigaam", "model_class": "ctc", "model_name": "v3_ctc", "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": "CTCHead", "head.feat_in": 768, "head.num_classes": 34, "decoding": "CTCGreedyDecoding", "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). Identical to the rnnt branch." }, "varying_across_variants": [ "model_class", "model_name", "head.type", "head.num_classes", "tokenizer.kind", "tokenizer.vocab_inline" ] }, "dtype": { "expected": "float32", "source": "manual", "evidence": "pytorch_model.bin is 441 MB ≈ 110M 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": 3.0, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_CTC row from the model card 5-domain WER summary." }, { "dataset": "Golos Farfield", "language": "ru", "metric": "wer", "score": 4.5, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_CTC row from the model card." }, { "dataset": "Natural Speech", "language": "ru", "metric": "wer", "score": 7.8, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_CTC row from the model card." }, { "dataset": "Callcenter", "language": "ru", "metric": "wer", "score": 10.3, "score_unit": "percent", "source": "https://huggingface.co/ai-sage/GigaAM-v3", "notes": "V3_CTC row from the model card." }, { "dataset": "Aggregate (11-set average)", "language": "ru", "metric": "wer", "score": 9.1, "score_unit": "percent", "source": "https://github.com/salute-developers/GigaAM/blob/main/evaluation.md", "notes": "V3_CTC 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_ctc')`.", "architecture_pattern": "encoder-ctc", "known_risks": [ "Encoder + frontend are expected IDENTICAL to gigaam-v3-e2e-rnnt and gigaam-v3-rnnt. Stage 3 must confirm encoder-tensor reuse. Only the head differs.", "CTCHead is `nn.Conv1d(768, num_classes=34, kernel_size=1)` + `log_softmax(-1)`. Blank id = 33 = num_classes - 1.", "Charwise tokenizer path identical to gigaam-v3-rnnt: vocab list inline in config.json, NO tokenizer.model. Detokenization is plain string join. The convert script must read the inline vocab and emit it into the GGUF; reuse the same charwise loader path as gigaam-v3-rnnt.", "Vocab: 33 entries — space + 32 lowercase Cyrillic letters (а-я, no ё). Lowercased Russian only, no English, no digits, no punctuation, no casing.", "Rotary positional embeddings inside Conformer attention — same risk as siblings.", "Conv1d subsampling, SpecScaler log, center=false, mel_scale=htk, mel_norm=null — same gotchas.", "Stage 7 FLEURS ru gate: same text-norm requirement as gigaam-v3-rnnt — strip punct, lowercase, fold ё→е on the FLEURS reference before WER computation.", "Long-form (>25s): out of scope for C++ parity (PyAnnote dependency)." ], "intake_gaps": [ { "field": "tokenizer.vocab_sha256", "reason": "Computed at Stage 2. Expected to MATCH gigaam-v3-rnnt's vocab_sha256 (same lowercased-Russian charwise vocab) — confirms vocab reuse across the two non-e2e branches." }, { "field": "dtype.expected_f32_tensors", "reason": "Filled at Stage 4." }, { "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_ctc').transcribe()` over `fleurs:ru` and use the measured number as the upstream anchor." } ] }