card: remove the Training (v2) section

README.md

@@ -146,21 +146,6 @@ sf.write("restored48k.wav", y, out_sr)
 | `examples/` | five public CallHome calls (TalkBank), each one channel with both parties: the input, and the v2 and v1 restorations |
 | `assets/` | the figures on this card |
 
-## Training (v2)
-
-- **Data.** Clean speech at 44.1 kHz or more from public corpora and in-house studio recordings, about 2,760 hours.
-- **Degradation, online.** Telephone channel, codecs and codec chains, line and background noise, cross-talk from other speakers,
-  packet loss, handset colouring and silence comfort noise. The noise and level settings are fitted to real call-centre recordings,
-  and every training pair stays recoverable (raised noise floors, at most two acoustic effects stacked).
-- **Schedule.** 100,000 steps from scratch at an effective batch of 12 six-second windows. Then 30,000 steps on the files whose
-  energy reaches 18 kHz (1.15 million files), with an extra mel loss on the band above 8 kHz.
-  Then 20,000 steps in which a fifth of the windows hold two parties in one channel, with both parties in the target.
-  These windows come from 20,000 synthetic two-party calls between studio speakers. The 16,828 calls without an in-house voice
-  are public: [Scicom-intl/Synthetic-Stereo-Calls](https://huggingface.co/datasets/Scicom-intl/Synthetic-Stereo-Calls).
-- **Losses.** Multi-resolution mel, GAN (adversarial and feature matching), F0, multi-resolution STFT and complex STFT, WavLM
-  features, and an energy loss that charges dropped speech eight times more than added energy. The encoder's LoRA adapter trains
-  together with the decoder from the first step.
-
 ## Limitations
 
 - The benchmarks are one operator's calls in English, Malay and Mandarin. Other lines and languages may differ.