Open Post-Training Recipes
Openly released data, code and methods (instruction tuning, preference optimisation, verifiable-reward RL) that turn base models into assistants.
Tier Grades
Tier A: Frontier AI labs
feasible
Ai2, Hugging Face and Google publish the recipes.
Tier B: Indonesian AI labs
feasible
Open recipes (Tulu 3) let an Indonesian lab post-train an open base without inventing methods.
Tier C: Institution units such as ITB
feasible
Applying a released recipe to an 8B base is institution-scale.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Delivery teams can apply SFT and DPO recipes for client-specific assistants.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeTulu 3 builds on Llama 3.1 base models; compute is not in the abstract.
- DataTulu 3 releases its data, code and recipes, with SFT, DPO and Reinforcement Learning with Verifiable Rewards (RLVR); it beats the instruct versions of Llama 3.1, Qwen 2.5 and Mistral.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 23 authors on Tulu 3.
- Time horizonMonths.
- Shelf-life signalsFully open releases (OLMo, Tulu) with data and code, RLVR moving from labs to open recipes
- Pipeline roleTier A and open labs release recipes; Tier B applies them to Indonesian models; Tier D tunes client assistants.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
TuluLLMexemplar
Zephyr 7B betaLLMexemplar
BLOOMZLLMexemplar
Flan-T5LLMexemplarOlmo-Hybrid-7BLLMproduced byOlmo 3.1 32B InstructLLMproduced byOlmo 3LLMproduced byOLMo 2 32B (OLMo-2-0325-32B-Instruct)LLMproduced byOLMo-2-0325-32BLLMproduced byOlmo 3 7BLLMproduced byLlama 3.1 Tulu 3 70BLLMproduced byLlama 3.1 Tulu 3 8BLLMproduced byOLMo 2LLMproduced byCendolLLMproduced by
Aya 101LLMproduced byOLMo 7B (first release)LLMproduced by
Neural-Chat 7B v3.3LLMproduced byTulu 2 DPO 70BLLMproduced by
Persimmon-8B-ChatLLMproduced byBLOOMZ 7B1LLMproduced byFlan-T5 baseLLMproduced byFlan-T5 largeLLMproduced byFlan-T5 XXLLLMproduced byOlmo-3-1125-32BLLMproduced byLinks
R&D TopicOpen Post-Training Recipes
ModelsTuluAllen AI, LLMBLOOMZBigScience Workshop, LLMFlan-T5Google, LLMZephyr 7B betaHugging Face, LLM+21 more
TuluAllen AI, LLMBLOOMZBigScience Workshop, LLMFlan-T5Google, LLMZephyr 7B betaHugging Face, LLM+21 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 25
Olmo 3Allen AI, LLMproduced byCendolIndoNLP, LLMproduced byOlmo 3.1 32B InstructAllen AI, LLMproduced byOlmo-3-1125-32BAllen AI, LLMproduced byOlmo-Hybrid-7BAllen AI, LLMproduced byFlan-T5Google, LLMexemplarOLMo 2Allen AI, LLMproduced byOLMo 2 32B (OLMo-2-0325-32B-Instruct)Allen AI, LLMproduced byOlmo 3 7BAllen AI, LLMproduced byOLMo-2-0325-32BAllen AI, LLMproduced byTuluAllen AI, LLMexemplarBLOOMZBigScience Workshop, LLMexemplar+13
BLOOMZ 7B1BigScience Workshop, LLMproduced byFlan-T5 baseGoogle, LLMproduced byAya 101Cohere Labs, LLMproduced byFlan-T5 largeGoogle, LLMproduced byFlan-T5 XXLGoogle, LLMproduced byLlama 3.1 Tulu 3 70BAllen AI, LLMproduced byLlama 3.1 Tulu 3 8BAllen AI, LLMproduced byOLMo 7B (first release)Allen AI, LLMproduced byTulu 2 DPO 70BAllen AI, LLMproduced by
Vicuna-7B-v1.5LMSYS, LLMproduced byZephyr 7B betaHugging Face, LLMexemplarNeural-Chat 7B v3.3Intel, LLMproduced byPersimmon-8B-ChatAdept AI, LLMproduced byPapers
- Training language models to follow instructions with human feedbackLong Ouyang et al., arXiv preprint, 2022 founding
- Tulu 3: Pushing Frontiers in Open Language Model Post-TrainingNathan Lambert et al., arXiv preprint, 2024 state of the art
- Direct Preference Optimization: Your Language Model is Secretly a Reward ModelRafael Rafailov et al., arXiv preprint, 2023 state of the art
- Zephyr: Direct Distillation of LM AlignmentLewis Tunstall et al., arXiv preprint, 2023 state of the art
- Open Problems and Fundamental Limitations of Reinforcement Learning from Human FeedbackStephen Casper et al., arXiv preprint, 2023 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2411.15124 https://arxiv.org/abs/2411.15124
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Open Post-Training Recipes", https://aurion.semai.id/topics/open-post-training-recipes/, records of October 2026.