zai-org/ChatGLM-6B

This system is a repository for the ChatGLM-6B large language model, providing tools for local interaction, API access, and multi-GPU deployment. It supports model refinement through a user feedback program for bad cases and offers a P-Tuning v2 workflow for efficient fine-tuning with reduced memory requirements.

The repository is initialized with the ChatGLM-6B model code, including an API server (api.py) for programmatic access and multiple demo interfaces (web\_demo.py, cli\_demo.py, and vision demos) for local interaction. A key feature added is multi-GPU deployment support via utils.py, which automatically configures device mapping to distribute the model across multiple GPUs, addressing cross-device tensor placement errors on Linux. The release also includes comprehensive documentation (README, FAQ, PROJECT.md) and licensing files (Apache 2.0 for code, specific license for model weights).

(repo-wide) · high confidence

This change adds a complete P-Tuning v2 implementation for the ChatGLM-6B model, allowing users to fine-tune the model with significantly reduced memory requirements (as low as 7GB) by optimizing only 0.1% of the parameters via a PrefixEncoder. The update includes training scripts (\train.sh\, \train\_chat.sh\) for both standard generation and multi-turn dialogue datasets, evaluation scripts (\evaluate.sh\, \evaluate\_finetune.sh\), and a Gradio-based web demo (\web\_demo.py\) that supports loading P-Tuning v2 checkpoints. It also provides configuration arguments for quantization, soft prompt length, and history handling, alongside documentation detailing the usage and performance metrics compared to full fine-tuning and LoRA.

ptuning · high confidence

The project introduces a new 'improve' directory containing documentation and sample data for the ChatGLM-6B Badcase Feedback Program. This initiative invites users to submit examples of poor model performance (bad cases) along with corrected responses to help refine the model. The included \README.md\ explains the program's goals, data submission guidelines, and privacy policies, while \data\_sample.jsonl\ provides example JSON lines showing the expected format for user contributions (prompt and response pairs).

improve · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• CAI 43 → 38 (-5.4)
• Rubric changed (rubric-2026.08.15 → rubric-2026.09.15) — scores are not directly comparable.

• Code Health 83 → 73 (-10.1)
• Architecture 94 → 99 (+5.3)
• Maturity 44 → 39 (-5.3)
• Readiness 17 → 11 (-5.8)
• Security 100 → 97 (-2.9)

• Dimension evaluation failed
• Duplicated block (11 lines × 2) (ptuning/trainer.py)
• Duplicated block (14 lines × 2) (ptuning/trainer.py)
• Duplicated block (5 lines × 2) (ptuning/trainer.py)
• Duplicated block (7 lines × 2) (ptuning/trainer_seq2seq.py)
• Duplicated block (9 lines × 2) (ptuning/trainer_seq2seq.py)
• No exposed public API
• early-stage repository — too little history to judge knowledge freshness
• main.main (cognitive 63) (main)
• main.main (cyclomatic 42) (main)
• single-commit history — no usable git history window to measure hotspots
• single-maintainer — knowledge-concentration (bus factor) risk

• Banned license: mdtex2html
• Critical CVE: [GHSA redacted] (requirements.txt)
• Documentation: no licence statement (README.md)
• Duplicated block (10–11 lines × 3) (ptuning/trainer.py)
• Duplicated block (11 lines × 3) (ptuning/main.py)
• Duplicated block (12–13 lines × 2) (ptuning/trainer.py)
• Duplicated block (13–15 lines × 2) (ptuning/trainer.py)
• Duplicated block (15–16 lines × 2) (ptuning/main.py)
• Duplicated block (17 lines × 2) (ptuning/trainer.py)
• Duplicated block (20 lines × 3) (ptuning/web_demo.py)
• Duplicated block (23–24 lines × 2) (ptuning/trainer.py)
• Duplicated block (39–47 lines × 2) (ptuning/trainer_seq2seq.py)
• Duplicated block (5 lines × 2) (ptuning/trainer.py)
• Duplicated block (5 lines × 2) (ptuning/trainer_seq2seq.py)
• Duplicated block (5 lines × 2) (ptuning/web_demo.py)
• Duplicated block (5 lines × 2) (web_demo_vision.py)
• Duplicated block (8 lines × 3) (ptuning/web_demo.py)
• Duplicated block (8 lines × 3) (ptuning/web_demo.py)
• Duplicated block (9 lines × 2) (ptuning/trainer.py)
• FileTooLong: ptuning/trainer.py (ptuning/trainer.py)
• …and 44 more

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

zai-org/ChatGLM-6B was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.