License Incompatibility
Hi, I’d like to report a potential license conflict in THUDM/cogvlm2-llama3-caption
. This model appears to be fine-tuned frommeta-llama/Llama-3.1-8B-Instruct
, which is governed by the LLaMA 3.1 Community License.
However, the downstream model is currently being distributed under the CogVLM2 LICENSE, which appears to conflict with key terms in the original Meta license — especially around redistribution rights, commercial usage, and field-of-use restrictions.
⚠️ Key violations of META LLAMA 3.1 COMMUNITY LICENSE AGREEMENT:
Clause 1.b – Redistribution and Use:
• Derivative models must retain the original LLaMA 3.1 license
• No sublicensing under a different or conflicting license
• Must include full license text and usage restrictions from Meta
Clause 1.c – Acceptable Use:
• All downstream use must adhere to Meta’s Acceptable Use Policy
• Redistributors must not remove or weaken these obligations
Clause 2 – No Additional Grant:
• No party may add extra permissions or usage terms beyond Meta’s license
Clause 3 – Commercial Restrictions:
• Any commercial use (especially beyond 700M MAU) requires a separate license from Meta
Conflicting Clauses from the CogVLM2 LICENSE:
• Allows commercial use after registration, without Meta’s explicit approval
• Adds geopolitical restrictions (e.g., “must not harm national unity”)
• Includes undefined societal restrictions (e.g., “must not harm public interest”)
• Replaces the LLaMA 3.1 license with its own terms, without mention of Meta’s license
🔹 Suggestion:
1. Add this NOTICE to the model card or repo:
> “This model is based on LLaMA 3.1 and is licensed under the LLaMA 3.1 Community License, Copyright © Meta Platforms, Inc.”
2. Remove CogVLM2 LICENSE or clearly separate it only for components unrelated to LLaMA
3. Include a clear note that any use must comply with Meta’s Acceptable Use Policy and commercial limitations
4. Replace the current license with the full LLaMA 3.1 Community License
Let me know if I’ve misunderstood the provenance of this model — happy to help revise this if needed!
Thanks for your attention!
Would love to hear your view on this!
CogVLM2 is available in variants trained on different base models. In our GitHub repository (https://github.com/THUDM/CogVLM2), any version that uses Llama 3 as its foundation must comply with the Llama 3 requirements, taking the most stringent clauses from each of the two licenses.