Upload README.md with huggingface_hub

README.md

@@ -20,13 +20,14 @@ layout and stable URLs for phoonnx voice entries.
 | Path | Source | License |
 |---|---|---|
 | `0.6B/model.onnx` (+ `model.onnx_data`), `0.6B/tokenizer.json` | [OuteAI/OuteTTS-1.0-0.6B-ONNX](https://huggingface.co/OuteAI/OuteTTS-1.0-0.6B-ONNX) | Apache-2.0 |
-| `1B/model.onnx` (+ `model.onnx_data`), `1B/tokenizer.json` | [OuteAI/Llama-OuteTTS-1.0-1B-ONNX](https://huggingface.co/OuteAI/Llama-OuteTTS-1.0-1B-ONNX) | **CC-BY-NC-SA-4.0 (non-commercial)** |
+| `1B/model.onnx` (+ `model.onnx_data`) | **re-exported by phoonnx** from [OuteAI/Llama-OuteTTS-1.0-1B](https://huggingface.co/OuteAI/Llama-OuteTTS-1.0-1B) - see below | **CC-BY-NC-SA-4.0 (non-commercial)** |
+| `1B/tokenizer.json` | [OuteAI/Llama-OuteTTS-1.0-1B-ONNX](https://huggingface.co/OuteAI/Llama-OuteTTS-1.0-1B-ONNX) | **CC-BY-NC-SA-4.0 (non-commercial)** |
 | `dac/decoder_model.onnx`, `dac/encoder_model.onnx`, `dac/config.json` | [OuteAI/DAC-speech-v1.0-ONNX](https://huggingface.co/OuteAI/DAC-speech-v1.0-ONNX) (weights: [ibm-research/DAC.speech.v1.0](https://huggingface.co/ibm-research/DAC.speech.v1.0)) | CDLA-Permissive-2.0 |
 | `speakers/en-female-1-neutral.json` | the `outetts` Python package (`version/v3/default_speakers`) | Apache-2.0 |
 
-The two model sizes carry **different licenses**. The 0.6B is Apache-2.0 and is what
-phoonnx uses by default. The 1B is CC-BY-NC-SA-4.0: it is **not licensed for commercial
-use**, and it is mirrored here only because it is the only member of the family that
+The two model sizes carry **different licenses**. The 0.6B is Apache-2.0 and covers 14
+languages; phoonnx resolves those 14 to it. The 1B is CC-BY-NC-SA-4.0 - **not licensed
+for commercial use** - and is here because it is the only member of the family that
 covers Arabic, Belarusian, Bengali, Lithuanian, Persian, Portuguese, Swahili, Tamil and
 Ukrainian. Check the 1B license before you deploy it.
 
@@ -52,16 +53,34 @@ Every export in this mirror was checked against the upstream torch weights on a
 | 0.6B `model_fp16.onnx` | 0.036 | 63/64 | no |
 | 0.6B `model_quantized` / `uint8` | 5.7 | 62/64 | no |
 | 0.6B `model_bnb4` / `int8` / `q4` / `q4f16` | 3.9 - 5.9 | 32-34/64 | no |
-| 1B `model.onnx` (fp32) | 12 | 61/64 | mirrored, **not indexed** |
+| 1B `model.onnx` - OuteAI's (fp32) | 12 | 61/64 | no, see below |
+| 1B `model.onnx` - this repo's re-export (fp32) | **1.78e-05** | **64/64** | **yes** |
 
 The 0.6B float32 export reproduces its torch weights exactly. Every quantized 0.6B export
 changes greedy decoding, so none is a safe default.
 
-The **1B float32 export does not reproduce its torch weights**: the logit correlation at
-the end of the prompt is 0.48 and greedy decoding diverges. The gap is present at short
-prompts too (0.08 max diff at 32 tokens) and grows with length, which points at the
-export rather than at accumulation. The 1B files are mirrored so the finding can be
-reproduced, but phoonnx does not index them.
+### The official 1B ONNX export is broken; this one is not
+
+`OuteAI/Llama-OuteTTS-1.0-1B-ONNX` **does not reproduce its own torch weights**, even in
+float32. On the prompt above its last-position logits are off by 12, the logit
+correlation is **0.48**, prefill argmax agrees on only 1580 of 1843 positions, and greedy
+decoding diverges from `OuteAI/Llama-OuteTTS-1.0-1B`. The error is already present on a
+32-token prompt (0.08 max diff) and never closes (correlation 0.96-0.98 at 512-2048
+tokens), so it is a property of the export, not of long-context accumulation. The 0.6B
+export on the same prompt differs by 3.8e-05. Every quantized 1B export inherits the
+defect and adds to it.
+
+The `1B/model.onnx` in **this** repository is therefore **not** OuteAI's file. It was
+re-exported from the torch checkpoint with
+[`scripts/conversion/outetts/export_outetts_onnx.py`](https://github.com/TigreGotico/phoonnx/blob/dev/scripts/conversion/outetts/export_outetts_onnx.py)
+and verified the same way: **max abs logit diff 1.78e-05, mean 3.74e-06, logit
+correlation 1.00000000, greedy 64/64**, and 64/64 again under the windowed repetition
+penalty OuteTTS actually samples with. The same script reproduces the 0.6B to 1.6e-05,
+which is how it is checked against a graph already known to be good.
+
+One difference to know about: the re-exported graph returns `logits` shaped `[1, 1, V]` -
+the final row only - where OuteAI's exports return every position. Any consumer that
+reads the last row, as transformers.js and phoonnx both do, works with either.
 
 ## Use it from phoonnx
 
@@ -75,8 +94,9 @@ for chunk in voice.synthesize("Hello from OuteTTS running in phoonnx."):
     ...  # chunk.audio_float_array, 24 kHz mono
 ```
 
-Voice ids run `outetts/0.6B/<lang>` for the 14 trained languages, and every one of them
-resolves to this repository. The engine drives the
+Voice ids run `outetts/0.6B/<lang>` for the 14 languages the 0.6B was trained on and
+`outetts/1B/<lang>` for the nine only the 1B covers (ar, be, bn, fa, lt, pt, sw, ta, uk).
+Every one of them resolves to this repository. The engine drives the
 LM with a KV cache, then decodes the two DAC codebooks to a waveform.
 
 ## How it works