Baichuan-7B Repo Skill

Use this skill when a task involves the Baichuan-7B repository, its custom Transformers model code, official Hugging Face loading pattern, C-Eval/MMLU benchmark scripts, or DeepSpeed pretraining demo. The skill is self-contained for operating decisions; do not reopen the original README or source scripts unless the user's active task is explicitly to edit a checkout.

Route by task

User intent	Read	Why
Load Baichuan-7B, inspect BaiChuanConfig, debug generation/cache behavior, or run a tiny architecture smoke	architecture-and-loading	Owns model classes, config defaults, trust_remote_code loading, and safe local-source checks.
Prepare C-Eval or MMLU evaluation, validate benchmark layout, render commands, or interpret output files	evaluation-workflows	Owns Chinese/English benchmark preflights, dataset layout, output artifacts, and benchmark-specific failures.
Prepare pretraining data, validate tokenizer.model, DeepSpeed JSON/hostfile, render launch commands, or explain checkpoints	pretraining-and-deepspeed	Owns corpus sharding, tokenizer placement, DeepSpeed setup, and safe training dry-runs.
Check class signatures, dependency surfaces, or source-derived architecture facts shared across workflows	API reference	Summarizes verified public classes, signatures, model dimensions, and script surfaces.
Diagnose install/import/backend/model-weight problems before choosing a sub-skill	troubleshooting	Covers cross-cutting dependency pins, CUDA/xFormers, model assets, datasets, and training resource boundaries.
Decide whether this skill is stale for a checkout	repo provenance	Records the commit, dirty state, dependency pins, and evidence paths used to create this skill.

Operating assumptions

• Baichuan-7B is a decoder-only 7B causal language model for Chinese and English with a 4096-token training context, 64k vocabulary, RMSNorm, SwiGLU MLP, rotary embeddings, and LLaMA-like design choices.
• The repository has no setup.py or pyproject.toml package metadata in this snapshot. Local source checks therefore use a Baichuan checkout path or a compatible model directory rather than a pip distribution import.
• The official README loading path uses AutoTokenizer.from_pretrained(..., trust_remote_code=True) and AutoModelForCausalLM.from_pretrained(..., device_map="auto", trust_remote_code=True).
• Full 7B generation, benchmark scoring, and DeepSpeed training require model weights/tokenizer files, CUDA-capable runtime, and task-specific data. The bundled helpers are safe preflight/smoke tools and do not download weights, fetch datasets, or launch training by default.
• Code is Apache-2.0 licensed. Model weights have a separate Baichuan model license; check the model source before commercial or redistribution decisions.

Setup checklist for a later runtime

1. Choose the workflow route above before installing optional packages.
2. Install the repo-documented stack when feasible: deepspeed==0.9.2, numpy==1.23.5, sentencepiece==0.1.97, torch==2.0.0, transformers==4.29.1, and xformers==0.0.20. If newer CUDA/Torch wheels are required, treat compatibility as a runtime decision and run the safe checks again.
3. For architecture-only checks, run the tiny smoke from the architecture sub-skill; it needs no official weights or datasets.
4. For real inference or evaluation, ensure the model identifier or local model directory contains compatible weights, config, tokenizer assets, and trusted remote-code files.
5. For benchmark runs, validate C-Eval or MMLU prerequisites with the evaluation preflight helper before loading the model.
6. For pretraining, validate corpus shards, tokenizer.model, DeepSpeed config, hostfile, and checkpoint path before rendering or launching any command.

Safe first checks

Architecture-only smoke against an active Baichuan-style checkout:

python sub-skills/architecture-and-loading/scripts/local_model_smoke.py --repo-root /path/to/Baichuan-7B

Benchmark preflight without executing inference:

python sub-skills/evaluation-workflows/scripts/check_evaluation_inputs.py ceval --model /path/to/model --shot 5 --split val

Training preflight without launching DeepSpeed:

python sub-skills/pretraining-and-deepspeed/scripts/validate_training_inputs.py \
  --data-dir /path/to/data_dir \
  --tokenizer-path /path/to/tokenizer.model \
  --deepspeed-config /path/to/deepspeed.json \
  --hostfile /path/to/hostfile

Stop conditions

Stop and ask for explicit user resources before:

• downloading official weights, benchmark datasets, or external benchmark checkouts;
• running full C-Eval/MMLU inference;
• launching DeepSpeed or distributed training;
• changing pinned package versions in a user-owned environment;
• accepting unverified CUDA/DeepSpeed behavior as a completed benchmark or training result.