File size: 1,554 Bytes
2bc4826
 
 
9b5a824
2bc4826
9b5a824
 
 
 
 
 
 
 
 
2bc4826
 
9b5a824
 
2bc4826
9b5a824
 
 
 
 
 
 
 
 
 
 
 
 
2bc4826
9b5a824
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
---
dataset_info:
  features:
    - name: source
      dtype: string
    - name: messages
      list:
        - name: role
          dtype: string
        - name: content
          dtype: string
        - name: thinking
          dtype: string
    - name: text
      dtype: string
  splits:
    - name: train
      num_examples: 84764
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - text-generation
language:
  - en
tags:
  - minicpm-5
  - reasoning
  - uncensored
  - chatml
---

# Sanitized SFT Dataset (85k)

This dataset contains exactly 84,764 rows optimized for **MiniCPM-5** Supervised Fine-Tuning (SFT) and Importance Matrix (Imatrix) calibration.

## Composition
- **Uncensored (28k):** High-quality, zero-refusal conversational data pulled from Alpaca-Uncensored and ShareGPT unfiltered variants.
- **Reasoning & Factual (57k):** Logic, math, and code-heavy data pulled from DeepSeek-v4, Claude 3.5, MetaMath, Hunter-Alpha, and Kimi-Cyber-Reasoning.

## Formatting
- The `text` column contains pre-rendered strings using the strict **MiniCPM-5 Jinja ChatML format**.
- **Sanitized:** Internal literal tags (like `<|im_end|>` or `<think>` discussed in coding problems) have been escaped to prevent tokenizer poisoning during training.
- Native `<think>` blocks are enforced for every assistant turn (populated if reasoning exists, explicitly empty if not).

## Usage
Perfect for direct loss-masking during SFT or passing to `llama.cpp` for accurate quantization calibration.