General Reasoning

End-to-end pipeline: synthetic data generation → SFT filtering → training → inference → evaluation.

---

Models and Data

Type	Name	Hugging Face
Model	LightThinker-Plus-Qwen	zjunlp/LightThinker-Plus-Qwen
Model	LightThinker-Plus-Llama	zjunlp/LightThinker-Plus-Llama
SFT Data	General Reasoning	zjunlp/LightThinker-Plus-GeneralReasoning-SFT

---

Environment Setup

Requirements: Python 3.12, CUDA 12.x, PyTorch 2.7

Step 1 — Create and activate the conda environment:

conda env create -f environment.yml
conda activate lt_plus_general_reasoning

The environment.yml uses Tsinghua mirrors. If you are outside China, replace the channels section with - defaults or - conda-forge.

Step 2 — Install flash_attn via pre-built wheel:

flash_attn is not included in environment.yml and must be installed separately. Download the wheel matching your CUDA / PyTorch / Python versions from:

https://github.com/Dao-AILab/flash-attention/releases

The wheel used in development:

# CUDA 12.x, PyTorch 2.7, Python 3.12, CXX11 ABI=TRUE
pip install flash_attn-2.8.3+cu12torch2.7cxx11abiTRUE-cp312-cp312-linux_x86_64.whl

Step 3 - Set environment variables

Copy .env.example to the repository root as .env, then modify it with your credentials and settings. The scripts auto-load <repo_root>/.env.

cp .env.example ../.env
set -a && source ../.env && set +a

---

Raw Synthetic Data Generation

Configure credentials in <repo_root>/.env (auto-loaded by the script), set an input JSONL dataset, then run:

DATASET_PATH=/path/to/input.jsonl bash scripts/synthetic.sh

The script shards the dataset across NUM_PROCESSES parallel workers and runs synthetic/main.py on each shard.

Required .env / env vars:

Env var	Description
GR_SYN_MODEL_NAME	Model name served by the API endpoint
GR_SYN_BASE_URL	OpenAI-compatible API base URL
GR_SYN_API_KEY	API key

Required runtime env vars:

Env var	Description
DATASET_PATH	Input JSONL dataset

Optional env vars (with defaults):

Env var	Default	Description
OUTPUT_ROOT	results/	Root directory for per-process output
OUTPUT_BASE_NAME	run_<timestamp>	Base name for output files
SYSTEM_PROMPT	system_prompt_math_unified	Key into PROMPTS_MAP
TOOLSETS	TOOLS_UNIFIED_FINAL_1229	Key into TOOL_MAP
MAX_WORKERS	20	Threads per process
NUM_PROCESSES	4	Number of parallel processes

Output is written to $OUTPUT_ROOT/proc_<i>/ for each shard.

---

Filtering and Formatting for SFT

Convert raw synthetic trajectories into LLaMA-Factory-compatible training data:

python synthetic/data_curation/train_data_construct.py \
  --input  <path/to/raw_results.jsonl_or_dir> \
  --output synthetic/curated_data/ \
  --filtering \
  --mode both \
  --strict

Arguments:

Argument	Default	Description
--input	(required)	Input JSON/JSONL file or directory
--output	(required)	Output directory
--mode	both	Output format: stepwise, full, or both
--filtering	off	Enable quality filtering (length, correctness checks)
--strict	off	Drop entire trajectory on duplicate/near-duplicate reasoning steps (default: drop only the step)
--add_think_tag	off	Prepend <tool_call> tag to commit/expand/fold reasoning
--original_file	None	Original JSONL used for ID-level deduplication filtering

---

Training

Edit MODEL_PATH, TRAIN_DATA_PATH, and SAVE_PATH in scripts/train.sh, then run from ms-swift/:

# Make sure you are in the ms-swift/ directory
bash scripts/train.sh

The script runs swift sft with the following fixed settings:

Setting	Value
GPUs	8 (CUDA_VISIBLE_DEVICES 0–7)
Parallelism	ZeRO-3 DeepSpeed
Dtype	bfloat16 + Flash Attention
Train type	Full fine-tune
Max sequence length	16384
Epochs	3
Batch size	4 per device
Gradient accumulation	1
Warmup	5%
Save strategy	Per epoch (keep last 3)

Learning rate guidance:

• 1e-5 — vanilla CoT (--inference_type vanilla_cot)
• 2e-5 — LThinker interactive / other modes

---

Inference

Run via scripts/infer.sh. The script is controlled by environment variables:

Env var	Required	Default	Description
MODEL	yes	—	Path to the model directory
MODEL_TAG	no	basename $MODEL	Tag used in output directory naming
RESULT_ROOT	no	<repo>/results/eval	Root directory for all outputs
INFER_MODE	no	vanilla	vanilla or interactive
TOKENS_MODE	no	throughput	throughput or budget (interactive only)
BATCH_SIZE	no	32	Inference batch size

Inference modes:

Mode	Underlying args
vanilla	--inference_type vanilla_cot --system_prompt
interactive + throughput	--inference_type interactive --turn_max_tokens 1024 --max_tokens_setting throughput
interactive + budget	--inference_type interactive --max_tokens_setting budget

Usage:

# Vanilla CoT (default)
MODEL=/path/to/model bash scripts/infer.sh <GPU_ID> gsm8k

# Multiple datasets (comma-separated)
MODEL=/path/to/model bash scripts/infer.sh 0 --datasets gsm8k,math,gpqa

# Interactive mode (throughput)
INFER_MODE=interactive MODEL=/path/to/model bash scripts/infer.sh 0 gsm8k

# Interactive mode (budget)
INFER_MODE=interactive TOKENS_MODE=budget MODEL=/path/to/model bash scripts/infer.sh 0 gsm8k

# Custom output directory
INFER_MODE=interactive RESULT_ROOT=/data/results MODEL=/path/to/model bash scripts/infer.sh 0 gpqa

Results are saved under $RESULT_ROOT/$MODEL_TAG/$DATASET/<mode>_turn<N>_temp<T>_.../{run.log, ...}.

---

Evaluation

Aggregate inference results into a leaderboard and compute accuracy, peak tokens, dependency, and throughput:

python scripts/eval.py \
  --input_dir  path/to/results/ \
  --datasets   gsm8k,mmlu,gpqa,bbh \
  --dependency_mode ltthinker

Arguments:

Argument	Default	Description
--input_dir	(required)	Directory tree containing *results*.json files
--datasets	gsm8k,mmlu,gpqa,bbh	Comma-separated list of datasets to evaluate
--dependency_mode	ltthinker	Computation mode for dependency metric: vanilla or ltthinker
--max_repeats	3	Max result files to aggregate per model–dataset pair
--allow_incomplete	off	Skip strict sample-count check (useful for partial runs)
--show_files	off	Print per-file load details
--out_leaderboard_json	<input_dir>/leaderboard.json	Path for the aggregated leaderboard JSON
--out_wrong_json	<input_dir>/wrong_cases.json	Path for the wrong-case dump

Supported datasets and expected sample counts:

Dataset	Expected count
gsm8k	1319
mmlu	1027
gpqa	198
bbh	495

Files with a count mismatch are skipped unless --allow_incomplete is set.

Output columns (printed as a tab-separated table and saved to leaderboard.json):

Model_Config · AVG_Acc · AVG_Time(h) · AVG_Sec/Samp · AVG_Peak · AVG_Dep(M) · AVG_Turn · AVG_GenTok and per-dataset variants of each metric.