FLCgeigeigei/flcmind-datasets

FLCMind Datasets

Training data for FLCMind, a minimal-but-runnable LLM project built from scratch that covers the full pipeline: pretraining → SFT fine-tuning → deployment. The project implements a Llama-style architecture (RMSNorm / RoPE / GQA / SwiGLU) with optional MoE support.

This repository ships both training stages as 7z archives. Unzip them into the project's dataset/ directory and the training scripts can read them directly.

Files

File	Archive	After extraction	Purpose
pretrain_t2t_mini.7z	~265 MB	pretrain_t2t_mini.jsonl (~1.2 GB)	Pretraining: language modeling / text continuation
sft_t2t_mini.7z	~307 MB	sft_t2t_mini.jsonl (~1.7 GB)	SFT: instruction tuning / multi-turn chat

Only the archives are hosted here to save bandwidth and storage. As a result, the Hugging Face Dataset Viewer cannot preview the contents — download and extract them instead.

Quick start

# 1. Download (or click the files on the web page)
hf download FLCgeigeigei/flcmind-datasets --repo-type dataset --local-dir ./dataset

# 2. Extract (7-Zip required)
7z x dataset/pretrain_t2t_mini.7z
7z x dataset/sft_t2t_mini.7z

Data format

Pretraining — pretrain_t2t_mini.jsonl

One JSON object per line, with a single text field:

{"text": "A continuous passage used for language modeling..."}

SFT — sft_t2t_mini.jsonl

One JSON object per line, in multi-turn conversation format:

{
  "conversations": [
    {"role": "system", "content": "You are flcmind, a small but useful language model."},
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi! How can I help you?", "reasoning_content": "", "tools": "", "tool_calls": ""}
  ]
}

Field	Description
role	system / user / assistant / tool
content	Message content
reasoning_content	Optional reasoning trace
tools	Optional tool definitions (JSON string)
tool_calls	Optional tool calls (JSON string)

Loading

from datasets import load_dataset

# Pretraining
ds = load_dataset("json", data_files="pretrain_t2t_mini.jsonl", split="train")
print(ds[0]["text"][:200])

# SFT
ds = load_dataset("json", data_files="sft_t2t_mini.jsonl", split="train")
print(ds[0]["conversations"][1]["content"])

Training pipeline

cd trainer
python train_pretrain.py                       # → out/pretrain_768.pth
python train_full_sft.py --from_weight pretrain  # → out/full_sft_768.pth

See the project's README.md for deployment and MoE options.

Notes

• The data contains both Chinese and English samples.
• For learning and research use only; please comply with the licenses of the original data sources.
• To preview samples online, extract the JSONL files and upload them (or convert them to Parquet). The repository size will grow to roughly 3 GB.