OpenIntelligenceNet commited on
Commit
9b5a824
·
verified ·
1 Parent(s): 2bc4826

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +41 -19
README.md CHANGED
@@ -1,27 +1,49 @@
1
  ---
2
  dataset_info:
3
  features:
4
- - name: source
5
- dtype: string
6
- - name: messages
7
- list:
8
- - name: content
9
  dtype: string
10
- - name: role
 
 
 
 
 
 
 
 
11
  dtype: string
12
- - name: thinking
13
- dtype: string
14
- - name: text
15
- dtype: string
16
  splits:
17
- - name: train
18
- num_bytes: 1168260677
19
- num_examples: 84764
20
- download_size: 869191998
21
- dataset_size: 1168260677
22
  configs:
23
- - config_name: default
24
- data_files:
25
- - split: train
26
- path: data/train-*
 
 
 
 
 
 
 
 
 
27
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  dataset_info:
3
  features:
4
+ - name: source
 
 
 
 
5
  dtype: string
6
+ - name: messages
7
+ list:
8
+ - name: role
9
+ dtype: string
10
+ - name: content
11
+ dtype: string
12
+ - name: thinking
13
+ dtype: string
14
+ - name: text
15
  dtype: string
 
 
 
 
16
  splits:
17
+ - name: train
18
+ num_examples: 84764
 
 
 
19
  configs:
20
+ - config_name: default
21
+ data_files:
22
+ - split: train
23
+ path: data/train-*
24
+ task_categories:
25
+ - text-generation
26
+ language:
27
+ - en
28
+ tags:
29
+ - minicpm-5
30
+ - reasoning
31
+ - uncensored
32
+ - chatml
33
  ---
34
+
35
+ # Sanitized SFT Dataset (85k)
36
+
37
+ This dataset contains exactly 84,764 rows optimized for **MiniCPM-5** Supervised Fine-Tuning (SFT) and Importance Matrix (Imatrix) calibration.
38
+
39
+ ## Composition
40
+ - **Uncensored (28k):** High-quality, zero-refusal conversational data pulled from Alpaca-Uncensored and ShareGPT unfiltered variants.
41
+ - **Reasoning & Factual (57k):** Logic, math, and code-heavy data pulled from DeepSeek-v4, Claude 3.5, MetaMath, Hunter-Alpha, and Kimi-Cyber-Reasoning.
42
+
43
+ ## Formatting
44
+ - The `text` column contains pre-rendered strings using the strict **MiniCPM-5 Jinja ChatML format**.
45
+ - **Sanitized:** Internal literal tags (like `<|im_end|>` or `<think>` discussed in coding problems) have been escaped to prevent tokenizer poisoning during training.
46
+ - Native `<think>` blocks are enforced for every assistant turn (populated if reasoning exists, explicitly empty if not).
47
+
48
+ ## Usage
49
+ Perfect for direct loss-masking during SFT or passing to `llama.cpp` for accurate quantization calibration.