Update README.md

README.md

@@ -1,73 +1,269 @@
 ---
-library_name: transformers
-license: other
-base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
+language:
+- am
+- ar
+- bn
+- zh
+- cs
+- nl
+- en
+- fr
+- de
+- el
+- ha
+- he
+- hi
+- id
+- it
+- ja
+- jv
+- km
+- ko
+- lo
+- ms
+- mr
+- fa
+- pl
+- pt
+- ro
+- ru
+- es
+- sw
+- sv
+- tl
+- ta
+- te
+- th
+- tr
+- uk
+- ur
+- vi
+license: apache-2.0
+datasets:
+- lightblue/reasoning-multilingual-R1-Llama-70B-train
 tags:
-- llama-factory
-- full
-- generated_from_trainer
-model-index:
-- name: reasoning-multilingual-R1-Llama-70B-train
-  results: []
+- reasoning
 ---
 
-<!-- This model card has been generated automatically according to the information the Trainer had access to. You
-should probably proofread and complete it, then remove this comment. -->
+# lightblue/DeepSeek-R1-Distill-Qwen-14B-Multilingual
 
-# reasoning-multilingual-R1-Llama-70B-train
+<div style="width: 100%; height: 160px; 
+            display: flex; align-items: center; 
+            justify-content: center; 
+            border: 8px solid black; 
+            font-size: 120px; font-weight: bold; 
+            text-align: center; 
+            color: #438db8,
+            font-family: 'Helvetica Neue', sans-serif;">
+  <span style="color: #438db8;">R1</span>
+  &nbsp;
+  <span style="color: blue;">m</span>
+  <span style="color: green;">u</span>
+  <span style="color: purple;">l</span>
+  <span style="color: yellow;">t</span>
+  <span style="color: pink;">i</span>
+  <span style="color: cyan;">l</span>
+  <span style="color: magenta;">i</span>
+  <span style="color: lime;">n</span>
+  <span style="color: teal;">g</span>
+</div>
 
-This model is a fine-tuned version of [deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) on the reasoning-multilingual-R1-Llama-70B-train dataset.
-It achieves the following results on the evaluation set:
-- Loss: 0.4441
+This is a Deepseek distill finetune trained on multilingual Chain-of-Thought (CoT).
+When this model is prompted in a language, it will both think and respond in that language, unlike the original R1 which will often think in either Chinese or English.
+This will make the outputs of these AIs more understandable and explainable to a wider audience.
+Hopefully this will be useful to the AI community, particularly those developing for languages aside from English and Chinese.
 
-## Model description
+This model is a multilingual fine-tuned version of [deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B).
 
-More information needed
+Other fine-tuned versions of this model can be found in [our collection, here](https://huggingface.co/collections/lightblue/r1-multilingual-679c890166ac0a84e83e38fa).
 
-## Intended uses & limitations
+This model was trained was trained using our [lightblue/reasoning-multilingual-R1-Llama-70B-train](https://huggingface.co/datasets/lightblue/reasoning-multilingual-R1-Llama-70B-train) dataset for ~10 minutes on the 8 x L20 instance ([ecs.gn8is-8x.32xlarge](https://www.alibabacloud.com/help/en/ecs/user-guide/gpu-accelerated-compute-optimized-and-vgpu-accelerated-instance-families-1)) on [Alibaba Cloud](https://www.alibabacloud.com/).
 
-More information needed
+# How to use
 
-## Training and evaluation data
+When using these models, we recommend using a sampling temperature of between 0.5-0.7, [as per the original distilled R1 models](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B#usage-recommendations).
 
-More information needed
+Additionally, we have observed that the model sometimes tends to repeat for more niche languages, so we also recommend setting `repetition_penalty` to 1.1, or higher if the model repeats itself when processing your prompts.
 
-## Training procedure
+We include scripts to use this model in vLLM:
 
-### Training hyperparameters
+<ul>
+  <li><b>vLLM</b>
 
-The following hyperparameters were used during training:
-- learning_rate: 1e-05
-- train_batch_size: 1
-- eval_batch_size: 1
-- seed: 42
-- distributed_type: multi-GPU
-- num_devices: 8
-- total_train_batch_size: 8
-- total_eval_batch_size: 8
-- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
-- lr_scheduler_type: cosine
-- lr_scheduler_warmup_ratio: 0.01
-- num_epochs: 1.0
+Install [vLLM](https://github.com/vllm-project/vllm/) using `pip install vllm`.
 
-### Training results
+<details open>
+  <summary>Show vLLM code</summary>
+  
+```python
+from vllm import LLM, SamplingParams
 
-| Training Loss | Epoch  | Step | Validation Loss |
-|:-------------:|:------:|:----:|:---------------:|
-| 0.4962        | 0.1019 | 11   | 0.5000          |
-| 0.5313        | 0.2037 | 22   | 0.4791          |
-| 0.4692        | 0.3056 | 33   | 0.4685          |
-| 0.3876        | 0.4074 | 44   | 0.4595          |
-| 0.4768        | 0.5093 | 55   | 0.4542          |
-| 0.4985        | 0.6111 | 66   | 0.4496          |
-| 0.4687        | 0.7130 | 77   | 0.4465          |
-| 0.4484        | 0.8148 | 88   | 0.4449          |
-| 0.4809        | 0.9167 | 99   | 0.4442          |
+llm = LLM(
+    model="lightblue/DeepSeek-R1-Distill-Qwen-7B-Multilingual",
+    max_model_len=8_000
+)
 
+sampling_params = SamplingParams(
+    temperature=0.5, 
+    max_tokens=8_000
+)
 
-### Framework versions
+prompts = [
+    """学校には1クラスにつき20人の生徒がおり、クラスは合計3つあります。
+学校全体では男子と女子がそれぞれ50%ずついます。
+1つ目のクラスには女子が15人、2つ目のクラスには女子が12人います。
+3つ目のクラスには何人の男子がいますか？"""
+]
 
-- Transformers 4.48.1
-- Pytorch 2.5.1+cu124
-- Datasets 3.1.0
-- Tokenizers 0.21.0
+conversations = [
+    [{"role": "user", "content": x}] for x in prompts
+]
+
+outputs = llm.chat(conversations, sampling_params=sampling_params)
+
+for output in outputs:
+    print(output.outputs[0].text)
+
+# <think>
+# まず、学校の総生徒数を算出します。各クラスに20人の生徒があり、クラスは3つあるため、総生徒数は60人です。
+
+# 次に、学校全体で男子と女子は同じ人数で分布しています。したがって、男子と女子各有30人。
+...
+# したがって、3つ目のクラスの男子数は20 - 3 = 17人です。
+# </think>
+
+# **解答：**
+
+# 学校の総生徒数を算出します。
+...
+# **最終的な答え：**
+# \[
+# \boxed{17}
+# \]
+```
+
+</details></li>
+</ul>	
+
+# Evaluation
+
+Through some quick evaluation of our own, we found this model can produce much correctly formatted and accurate results for higher resource languages, such as Japanese, English, German, than lower resource languages, such as Amharic or Lao.
+
+We did a **very** quick evaluation of 5 questions with each dataset (written by me and translated by GPT4o Mini) on the [lightblue/DeepSeek-R1-Distill-Qwen-7B-Multilingual](https://huggingface.co/lightblue/DeepSeek-R1-Distill-Qwen-7B-Multilingual) model, and we find that the model is able to fairly reliably output the correct answers and in the correct language for a large variety of languages:
+
+For this evaluation, a score of >=0.8 is good, as one of the questions was very hard. The language detection was done using [pycld2](https://pypi.org/project/pycld2/) so errors may occur with the correct language being mistaken for another one.
+
+| language        |   Has a correct think statement |   Has the think statement in the correct language |   Is the response in the correct language |   Is the answer correct |
+|:----------------|------------:|------------------------:|----------------------:|-------------:|
+| Amharic         |         0.2 |                     0   |                   0   |          0   |
+| Arabic          |         1   |                     0.8 |                   0.8 |          0.6 |
+| Bengali         |         1   |                     1   |                   1   |          0.2 |
+| Chinese         |         1   |                     1   |                   1   |          0.8 |
+| Czech           |         1   |                     1   |                   1   |          0.8 |
+| Dutch           |         1   |                     1   |                   1   |          0.8 |
+| English         |         1   |                     1   |                   1   |          0.8 |
+| French          |         1   |                     1   |                   1   |          0.8 |
+| German          |         1   |                     1   |                   1   |          0.8 |
+| Greek           |         1   |                     1   |                   1   |          0.6 |
+| Hausa           |         0.4 |                     0   |                   0   |          0   |
+| Hebrew          |         1   |                     0.8 |                   1   |          0.6 |
+| Hindi           |         1   |                     1   |                   1   |          0.8 |
+| Indonesian      |         1   |                     1   |                   1   |          0.8 |
+| Italian         |         1   |                     1   |                   1   |          0.8 |
+| Japanese        |         1   |                     1   |                   0.8 |          0.6 |
+| Javanese        |         0.8 |                     0.2 |                   0.2 |          0.6 |
+| Khmer           |         0.6 |                     0.6 |                   0.6 |          0   |
+| Korean          |         1   |                     1   |                   1   |          1   |
+| Lao             |         0.4 |                     0.4 |                   0.4 |          0   |
+| Malay           |         1   |                     0.4 |                   0.4 |          0.8 |
+| Marathi         |         0.6 |                     0.4 |                   0.6 |          0.2 |
+| Persian (Farsi) |         0.6 |                     None*   |                   None*   |          0.2 |
+| Polish          |         1   |                     1   |                   1   |          0.6 |
+| Portuguese      |         1   |                     1   |                   1   |          0.8 |
+| Romanian        |         1   |                     1   |                   1   |          0.8 |
+| Russian         |         1   |                     1   |                   1   |          0.8 |
+| Spanish         |         1   |                     1   |                   1   |          0.8 |
+| Swahili         |         0.4 |                     0.4 |                   0.4 |          0   |
+| Swedish         |         1   |                     1   |                   1   |          0.8 |
+| Tagalog         |         1   |                     1   |                   1   |          0.8 |
+| Tamil           |         0.8 |                     0.8 |                   0.8 |          0.2 |
+| Telugu          |         0.8 |                     0.6 |                   0.8 |          0   |
+| Thai            |         1   |                     1   |                   1   |          0.8 |
+| Turkish         |         1   |                     1   |                   1   |          0.8 |
+| Ukrainian       |         1   |                     1   |                   1   |          0.8 |
+| Urdu            |         1   |                     1   |                   1   |          0.6 |
+| Vietnamese      |         1   |                     1   |                   1   |          1   |
+
+* There was an error with Farsi detection (my own fault) so we do not report Farsi scores.
+
+The evaluation code for this can be found [here](https://drive.google.com/file/d/1P33GpqvKmHoZUsWqqBPXHTToN2W7MDRG/view?usp=sharing).
+
+# Training code
+
+```yaml
+### model
+model_name_or_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: /root/LLaMA-Factory/examples/deepspeed/ds_z3_config.json
+
+### dataset
+dataset: reasoning-multilingual-R1-Llama-70B-train
+template: qwen
+cutoff_len: 4096
+overwrite_cache: true
+preprocessing_num_workers: 16
+packing: true
+
+### output
+output_dir: /root/train_outputs/DeepSeek-R1-Distill-Qwen-14B/reasoning-multilingual-R1-Llama-70B-train
+logging_steps: 1
+save_steps: 0.99999
+plot_loss: true
+overwrite_output_dir: true
+
+### train
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 1
+learning_rate: 1.0e-5
+num_train_epochs: 1.0
+lr_scheduler_type: cosine
+warmup_ratio: 0.01
+bf16: true
+ddp_timeout: 180000000
+
+### eval
+val_size: 0.01
+per_device_eval_batch_size: 1
+eval_strategy: steps
+eval_steps: 0.1
+```
+
+```bash
+echo '{
+  "reasoning-multilingual-R1-Llama-70B-train": {
+    "hf_hub_url": "lightblue/reasoning-multilingual-R1-Llama-70B-train",
+    "formatting": "sharegpt"
+  }
+}' > /root/LLaMA-Factory/data/dataset_info.json
+
+# # 14B Llama
+cd /root/LLaMA-Factory && llamafactory-cli train /root/reasoning_multilingual_train_14B.yaml
+rm -r /root/train_outputs/DeepSeek-R1-Distill-Qwen-14B/reasoning-multilingual-R1-Llama-70B-train/checkpoint*
+huggingface-cli upload lightblue/DeepSeek-R1-Distill-Qwen-14B-Multilingual /root/train_outputs/DeepSeek-R1-Distill-Qwen-14B/reasoning-multilingual-R1-Llama-70B-train
+```
+
+# License
+
+We share this model with the Apache 2.0 license.
+
+# Developed by
+
+<a href="https://www.lightblue-tech.com">
+<img src="https://www.lightblue-tech.com/wp-content/uploads/2023/08/color_%E6%A8%AA%E5%9E%8B-1536x469.png" alt="Lightblue technology logo" width="400"/>
+</a>
+
+This model was trained by Peter Devine ([ptrdvn](https://huggingface.co/ptrdvn)) for Lightblue