初始化项目，由ModelHub XC社区提供模型

Model: mesolitica/Malaysian-orpheus-3b-0.1-ft Source: Original Platform
2026-05-28 15:34:15 +08:00
commit caebf3d1ba
11 changed files with 496 additions and 0 deletions
--- a/README.md
+++ b/README.md
@@ -0,0 +1,105 @@
+---
+library_name: transformers
+base_model:
+- canopylabs/orpheus-3b-0.1-ft
+---
+
+# Malaysian canopylabs/orpheus-3b-0.1-ft
+
+Finetune [canopylabs/orpheus-3b-0.1-ft](https://huggingface.co/canopylabs/orpheus-3b-0.1-ft) on standard Malay and minimal Mandarin.
+
+## Training session
+
+Finetune on [Mesolitica/TTS](https://huggingface.co/datasets/mesolitica/TTS) to make the model able to generate Malay voice with minimal Mandarin.
+  
+## How we train
+
+1. LoRA on `["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj", "embed_tokens", "lm_head"]`.
+2. 128 Rank with alpha 256, or alpha of 2.0, but during merging, we use 1.5 ratio.
+3. Multipacking with proper SDPA causal masking to prevent document contamination and also make sure proper position ids.
+4. Chunk CE loss to reduce memory.
+
+Wandb at https://wandb.ai/huseinzol05/malay-orpheus-3b-0.1-ft-lora-128/workspace?nw=nwuserhuseinzol05
+
+## Example
+
+Load the model,
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer
+from snac import SNAC
+import torch
+import IPython.display as ipd
+
+def redistribute_codes(row):
+    row_length = row.size(0)
+    new_length = (row_length // 7) * 7
+    trimmed_row = row[:new_length]
+    code_list = [t - 128266 for t in trimmed_row]
+    layer_1 = []
+    layer_2 = []
+    layer_3 = []
+    for i in range((len(code_list)+1)//7):
+        layer_1.append(code_list[7*i][None])
+        layer_2.append(code_list[7*i+1][None]-4096)
+        layer_3.append(code_list[7*i+2][None]-(2*4096))
+        layer_3.append(code_list[7*i+3][None]-(3*4096))
+        layer_2.append(code_list[7*i+4][None]-(4*4096))
+        layer_3.append(code_list[7*i+5][None]-(5*4096))
+        layer_3.append(code_list[7*i+6][None]-(6*4096))
+    
+    with torch.no_grad():
+
+        codes = [torch.concat(layer_1), 
+            torch.concat(layer_2), 
+            torch.concat(layer_3)]
+        
+        for i in range(len(codes)):
+            codes[i][codes[i] < 0] = 0
+            codes[i] = codes[i][None]
+        
+        audio_hat = snac_model.decode(codes)
+        return audio_hat.cpu()[0, 0]
+
+snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
+snac_model = snac_model.to("cuda")
+
+tokenizer = AutoTokenizer.from_pretrained('mesolitica/Malaysian-orpheus-3b-0.1-ft')
+model = AutoModelForCausalLM.from_pretrained(
+    'mesolitica/Malaysian-orpheus-3b-0.1-ft', torch_dtype = torch.bfloat16
+).cuda()
+
+speakers = [
+    'Husein',
+    'Shafiqah Idayu',
+    'Anwar Ibrahim',
+    'KP'
+]
+
+speaker = speakers[0]
+text = 'Nama saya Husein, saya tak suka nasi ayam dan tak suka mandi, Xiàn zài wǒ yǒu bing chilling Wǒ hěn xǐ huān bing chilling.'
+prompt = f'<custom_token_3><|begin_of_text|>{speaker}: {text}<|eot_id|><custom_token_4><custom_token_5><custom_token_1>'
+input_ids = tokenizer(prompt,add_special_tokens = False, return_tensors = 'pt').to('cuda')
+
+with torch.no_grad():
+    generated_ids = model.generate(
+      **input_ids,
+      max_new_tokens=1200,
+      do_sample=True,
+      temperature=0.9,
+      top_p=0.8,
+      repetition_penalty=1.1,
+      num_return_sequences=1,
+      eos_token_id=128258,
+    )
+
+row = generated_ids[0, input_ids['input_ids'].shape[1]:]
+y_ = redistribute_codes(row)
+ipd.Audio(y_, rate = 24000)
+```
+
+<audio controls src="https://cdn-uploads.huggingface.co/production/uploads/5e73316106936008a9ee6523/NIOtl7B6Myw1eBd5Lf76l.wav"></audio>
+
+## Source code
+
+Source code at https://github.com/mesolitica/malaya-speech/tree/master/session/orpheus