Access to this repository is granted by the TrustLLM project on request

Research-only artifact of the TrustLLM project (EU). Access requests are reviewed manually by project administrators; a request does not imply approval.

NOLMA-8B SFT (release check pending)

Supervised fine-tunes of NOLMA-8B (anneal B, step 102,400) with the V2-8B mix-B2 SFT recipe (6,144 steps, 8k context), trained on MareNostrum 5 in October 2026. Research use only; access is gated and reviewed by the TrustLLM project administrators.

Each pre-registered run has its step-6144 export on its own branch:

Branch	Run
`b2n-seed420000-step-6144`	B2-N, seed 420000
`b2n-seed420001-step-6144`	B2-N, seed 420001
`b2n-seed420002-step-6144`	B2-N, seed 420002
`lc10-seed420000-step-6144`	Arm L: long-context share 10 %
`replay-seed420000-step-6144`	Arm R: 7.5 % replay of the anneal-B pretraining mix

The pre-registered release check compares these runs with the three V2-8B SFT models. The release candidate goes to main after that check. Until then, no branch is a release.

The router expert biases are stored in float32. Load with trust_remote_code=True, for example AutoModelForCausalLM.from_pretrained("TrustLLMeu/nolma-8B-sft", revision="b2n-seed420000-step-6144", trust_remote_code=True).

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for TrustLLMeu/nolma-8B-sft

Base model

TrustLLMeu/nolma-8B

Finetuned

(1)

this model