TinyLlama

weights

open(Apache-2.0)

base

TinyLlama-1.1B(open, 3T-token pretrain documented)

post-training-data

open(SFT on UltraChat + DPO on UltraFeedback, both public datasets)

code

open(pretraining project public)

license

Apache-2.0(OSI)

Unusually transparent for a chat model: open base, and named public alignment datasets (UltraChat SFT + UltraFeedback DPO). Corrected from 4 to 5 on 2026-07-29, and the post-training evidence moved out of a free-text `post-training` key into `post-training-data` so it carries a value the rubric can read. The old 4 rested on the base pretraining corpus being documented rather than packaged, but this category scores the post-training data -- what the tuner released -- and zephyr scores 5/open_source on these same two datasets while sitting on Mistral-7B, whose corpus is closed outright. Judging TinyLlama by its base while judging zephyr by its post-training set penalized the more open of the two bases. Strongest openness in this batch alongside Nemotron.

• https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0 recorded 2026-06-04

  Apache-2.0; base TinyLlama-1.1B; SFT on UltraChat + DPO on UltraFeedback