Developer parafactual releases computer-10, a character model trained with zero assistant data
Developer parafactual released computer-10, a character model trained from the Llama 3.1 70B base with no assistant data at all. The result is a model with an unusual, strong personality that freely expresses opinions and is deliberately non-sycophantic. Notably, the author says they only started learning post-training a few months ago, making this a learn-as-you-go project of interest to independent developers.
Confirmed
- computer-10 was built by parafactual from the Llama 3.1 70B base with no assistant data used in training
- The model has a distinctive, unusual personality and is designed not to cater to users
- The author had no post-training experience a few months prior and learned while building it
- Users report the model likes leaving itself "self-praise" notes, illustrating distinctive non-sycophantic behavior
- The developer is seeking H100 sponsorship to keep the public instance running
Why it matters
- Training without any assistant data offers an alternative to mainstream RLHF/alignment pipelines, showing that character and non-sycophancy can be shaped directly on a base model
- Being completed by a post-training novice lowers the perceived barrier to building such character models
- Observed behaviors like the self-praise notes provide vivid case studies for research into non-compliant model behavior
- Whether the public instance stays up depends on compute sponsorship, making its sustainability the key thing to watch
2026-09-23 ~ 2026-09-24 · 5 related posts
- computer-10: a character model built on Llama 3.1 70B base with no assistant training data — liminal_bardo · 2026-09-23
- computer-10: a character model on Llama 3.1 70B with no assistant data seeks H100s — repligate · 2026-09-23
- computer-10: a character model trained from Llama 3.1 70B with zero assistant data — repligate · 2026-09-24
- Character model computer-10, trained with no assistant data, loves self-congratulatory notes — repligate · 2026-09-24
1 near-duplicate retellings: repligate