Tulu 3 8B
Fully open post-training recipe on Llama 3.1 8B: SFT, DPO, and RLVR.
About
Allen AI's open post-training package applied to Llama 3.1 8B: supervised fine-tuning, direct preference optimization, and a new reinforcement learning method with verifiable rewards (RLVR).
The model is only half the story — the release includes the datasets, training code, and evaluation tooling, so anyone can reproduce the recipe. The paper reports it beating Llama 3.1 Instruct, Qwen 2.5, Mistral, and even GPT-4o-mini and Claude 3.5 Haiku on their evaluation suite.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Surpasses Llama 3.1 Instruct, Qwen 2.5, Mistral, GPT-4o-mini, and Claude 3.5 Haiku on Ai2's multi-task post-training evaluation suite (paper arXiv:2411.15124).
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Widely cited post-training reference; 5.8k+ likes on the 8B checkpoint; Recipe superseded by Tulu 3.1 but the 8B remains the standard open post-training baseline; Data, code, and training recipes fully open; base model carries the Llama 3.1 Community License. Manual review: SFT+DPO+RLVR pipeline and benchmark claims verified against the arXiv paper. License: Llama 3.1 Community License (inherited from base model).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Widely cited post-training reference; 5.8k+ likes on the 8B checkpoint
- · Recipe superseded by Tulu 3.1 but the 8B remains the standard open post-training baseline
- · Data, code, and training recipes fully open; base model carries the Llama 3.1 Community License
- · manual review: SFT+DPO+RLVR pipeline and benchmark claims verified against the arXiv paper