Deep-Dive FTDD-02 — OLMo 2/3 + Tülu 3 (Ai2)

Course: Course 3 — LLM Fine-Tuning Masterclass Deep-Dive: FTDD-02 Duration: 45 minutes Level: Senior Engineer and above Prerequisites: FT00 (Steering Stack), FT02 (Open Spectrum), FTDD-01 (MiniCPM)

The fully-open comparison. Allen Institute's OLMo and Tülu are what "open" means when a research lab commits to releasing everything — weights, data, code, and the complete post-training recipe. The counterpoint to MiniCPM's product-and-edge orientation.


Learning Objectives

After this deep-dive, you will be able to:

  1. Distinguish Ai2's OLMo (the open base) from Tülu 3 (the open post-training recipe) and explain why both are needed for full reproducibility.
  2. Map the OLMo 2 → OLMo 3 progression and the base/instruct/think variants, citing arXiv:2501.00656 for OLMo 2 and arXiv:2411.15124 for Tülu 3.
  3. Trace the Tülu 3 post-training pipeline (SFT → DPO → RLVR) and place each stage on the FT00 steering stack.
  4. Defend why fully-open (weights + data + code + recipe) matters for reproducibility — and why "weights-only open" is not enough.
  5. Contrast the OpenBMB/MiniCPM and Ai2/OLMo-Tülu philosophies: both open-data, but MiniCPM is product/edge-oriented while OLMo/Tülu is research-oriented.

The Subject

The Allen Institute for AI (Ai2) is the canonical fully-open lab. Where OpenBMB (FTDD-01) builds small models for product and edge deployment, Ai2 builds models whose primary contribution is the openness itself — the demonstration that a frontier-scale model can be released with every component a researcher needs to reproduce, audit, and extend it. Two releases carry this banner: OLMo (the open base) and Tülu 3 (the open post-training recipe).

Metric Value
Lab Allen Institute for AI (Ai2)
License Apache-2.0 (weights, data, code)
Open base OLMo 2 (arXiv:2501.00656) · OLMo 3 (Nov 2025, 7B + 32B)
Open post-train Tülu 3 (arXiv:2411.15124), up to 405B
Pipeline repo allenai/open-instruct
OLMo 2 nickname "2 OLMo 2 Furious"
OLMo 3 variants base · instruct · think
Tülu 3 stages SFT → DPO → RLVR (verifiable rewards)

Ai2 matters for this course because it is the strict end of the open spectrum (FT02). Where MiniCPM is open-data and open-recipe, OLMo and Tülu push further: they release intermediate checkpoints, the eval harness, and the exact data mixture. A skilled person can rebuild the model from scratch; a security team can prove no training-time exfiltration occurred. This is the tier that IL5/IL6 and air-gapped environments increasingly require (FT22).


OLMo — The Open Base

OLMo 2 — "2 OLMo 2 Furious"

OLMo 2 (arXiv:2501.00656) is the open base that established the modern fully-open standard. The nickname — "2 OLMo 2 Furious" — is the community's nod to the paper's density and the team's pace. What OLMo 2 contributes is not a benchmark win; it is the release of every component the FT02 open-recipe tier promises:

The OLMo 2 paper is the canonical "fully open" citation this course returns to in FT02. When the course says "open-recipe means you can rebuild and prove it," OLMo 2 is the example. Every other open-recipe release (SmolLM3, MiniCPM) is measured against the OLMo standard.

OLMo 3 — the three-variant release

OLMo 3, released November 2025, comes in two sizes (7B and 32B) and — critically — three variants:

The three-variant release is significant because it makes the steering stack (FT00) concrete. The base is Layer 1. The instruct is base + Layer 3 (SFT + DPO). The think is instruct + Layer 3 (GRPO). A student can load all three from the same family and see what each steering stage changed — a pedagogical gift that a single-variant release cannot provide. This is why OLMo 3 is the model to reach for when teaching the difference between SFT, DPO, and GRPO: the same base, three steers, directly comparable.


Tülu 3 — The Open Post-Training Recipe

If OLMo is the open base, Tülu 3 (arXiv:2411.15124) is the open post-training stack. Tülu 3's contribution is the full, reproducible recipe for turning a base model into an aligned, capable instruct model — released as data, code, and configs, at scales up to 405B parameters.

The three-stage pipeline

Tülu 3 post-training is a three-stage pipeline, and each stage maps cleanly onto the FT00 steering stack:

BASE MODEL (Layer 1)
    │
    ▼
[1] SFT  ────────────────── Layer 3 (format, instruction-following)
    │   (Supervised Fine-Tuning on curated instruction data)
    ▼
[2] DPO  ────────────────── Layer 3 (preference alignment)
    │   (Direct Preference Optimization on preference pairs)
    ▼
[3] RLVR ────────────────── Layer 3 (reasoning, verifiable rewards)
    │   (Reinforcement Learning on Verifiable Rewards — math, code, etc.)
    ▼
TÜLU 3 INSTRUCT MODEL

Stage 1 — SFT. Supervised Fine-Tuning on a curated instruction dataset. This is the format-and-injection-following steer (FT12). Tülu 3 releases the SFT data mixture, so you can see exactly what instructions the model was steered toward.

Stage 2 — DPO. Direct Preference Optimization on preference pairs (FT13). This is the preference-alignment steer — "this response is better than that one." Tülu 3 releases the preference dataset and the DPO hyperparameters.

Stage 3 — RLVR. Reinforcement Learning on Verifiable Rewards (FT14). This is the reasoning steer — the model generates solutions to problems with verifiable answers (math, code, logic), and correct solutions are rewarded. RLVR is the technique behind the "reasoning model" wave, and Tülu 3 is the open implementation of it.

The load-bearing property: every stage's data, code, and config is released. You can rebuild Tülu 3 from an OLMo base. You can ablate a single stage (what if you skip RLVR?). You can swap in your own data at any stage. This is what "open post-training recipe" means in practice — not a paper describing the recipe, but the recipe itself, executable.

The open-instruct repo

The practical handle for all of this is allenai/open-instruct — the repository that contains the Tülu 3 training code, data preparation scripts, and configuration files. If you want to reproduce Tülu 3, or build a Tülu-style pipeline on your own base, open-instruct is where you start. It is the post-training analogue of OLMo's pretraining code: the executable recipe, not the description.


Why Fully-Open Matters for Reproducibility

The argument for fully-open (weights + data + code + recipe) is not ideological. It is the FT02 argument, sharpened: reproducibility is the property that lets you prove things about a model, and reproducibility requires every component.

Three things only fully-open releases give you:

  1. Reproducibility in the strict sense. You can rebuild the model from a pinned commit and get the same weights (modulo nondeterminism, which is itself documented). This is the no-silent-drift argument (FT02) — you can prove the model you validated is the model you are running.
  2. Ablation. You can remove or change a single component (a data source, a post-training stage) and measure the effect. This is how research advances — and it is impossible with a weights-only release, where the data and recipe are opaque.
  3. Trust in the supply chain. You can audit the training pipeline end-to-end and rule out hidden training-time exfiltration. This is the property air-gapped and classified environments require (FT22).

The Tülu 3 release is the clearest demonstration of (2) — ablation — in the open ecosystem. Because the full post-training recipe is released, a researcher can ask "what does RLVR actually contribute?" by training Tülu 3 with and without the RLVR stage, on the same base, with the same SFT and DPO. With a weights-only release, that question is unanswerable.


OpenBMB vs Ai2 — Two Open Philosophies

Both OpenBMB (MiniCPM) and Ai2 (OLMo/Tülu) are open-data and open-recipe. Both ship Apache-2.0. But they have different centers of gravity, and the difference matters when you choose a base.

Property OpenBMB / MiniCPM Ai2 / OLMo + Tülu
Orientation Product / edge Research
Model sizes Small (1B–4B) + multimodal Mid-to-large (7B–405B), text-focused
Post-training recipe Disclosed, productized Fully released, ablation-ready (Tülu 3)
Multi-variant releases Modality axis (text/vision/omni) Steer axis (base/instruct/think)
Default use case Ship a small model on edge hardware Reproduce, audit, research the stack
Course role The on-ramp base (FT00 lab) The fully-open reference (FT02, FT22)

The summary: MiniCPM is what you ship; OLMo/Tülu is what you study. If your goal is a small model on a phone or an edge device, MiniCPM's modality axis and small sizes win. If your goal is to understand, audit, or reproduce the full training-and-post-training stack — especially for a regulated deployment that demands proof — OLMo and Tülu are the reference.

This is why the course uses both. FT00's lab loads MiniCPM5-1B because it is cheap to iterate on. FT02's audit references OLMo 2 because it is the fully-open standard. And when the course needs to teach what post-training actually does (FT12 SFT, FT13 DPO, FT14 GRPO), OLMo 3's base/instruct/think variants and Tülu 3's released pipeline are the clearest teaching tools — the same base, each steer visible.


Anti-Patterns

Treating "fully-open" as "production-ready"

OLMo and Tülu are research artifacts first. They are fully open, auditable, and reproducible — but they are not always the most capable models at their size point. A closed or weights-only model may outperform them on benchmarks. Choose them for openness, auditability, and reproducibility, not because they top every leaderboard. Capability is FT03's concern.

Assuming OLMo 3 "think" = GPT-class reasoning

The think variant is a reasoning-tuned model (RLVR), but it is a 7B/32B model, not a frontier model. It demonstrates the RLVR technique (FT14) openly — which is its research contribution — but it will not match a closed frontier reasoning model on the hardest tasks. Use it to understand reasoning fine-tuning, not to deploy a frontier reasoner.

Skipping the open-instruct repo

Tülu 3's value is in the recipe, and the recipe lives in allenai/open-instruct. Reading the paper without reading the repo gives you the description; reading the repo gives you the executable pipeline. If you want to reproduce or adapt Tülu 3, the repo is the primary source.


Key Terms

Term Definition
Ai2 Allen Institute for AI; the canonical fully-open lab
OLMo 2 The open base that established the modern fully-open standard (arXiv:2501.00656); nicknamed "2 OLMo 2 Furious"
OLMo 3 The three-variant release (Nov 2025, 7B + 32B): base, instruct, think
Tülu 3 The open post-training recipe (arXiv:2411.15124); SFT → DPO → RLVR, up to 405B
RLVR Reinforcement Learning on Verifiable Rewards — the reasoning steer (Tülu 3 stage 3, FT14)
open-instruct allenai/open-instruct — the repo containing the Tülu 3 training code, data prep, and configs
Fully-open Weights + data + code + recipe + checkpoints + eval — the strictest tier (FT02)
base / instruct / think OLMo 3's three variants: pretrained, instruction-tuned, reasoning-tuned

Lab Exercise

See 07-lab-spec.md. The lab compares OLMo 3's base, instruct, and think variants on a reasoning task — making the three steering stages (none, SFT+DPO, +RLVR) directly visible on the same base, no fine-tuning required.


References

  1. OLMo 2 paper — arXiv:2501.00656; the fully-open base standard.
  2. Tülu 3 paper — arXiv:2411.15124; the open post-training recipe.
  3. open-instruct repoallenai/open-instruct; the executable Tülu 3 pipeline.
  4. OLMo project pageallenai.org/olmo; the family overview.
  5. Course 3 FT02 — the Open Spectrum; OLMo/Tülu as the open-recipe reference.
  6. Course 3 FT14 — GRPO and verifiable rewards; RLVR is the open implementation.
  7. Course 3 FTDD-01 — MiniCPM; the product/edge counterpoint to OLMo/Tülu's research orientation.
# Deep-Dive FTDD-02 — OLMo 2/3 + Tülu 3 (Ai2)

**Course**: Course 3 — LLM Fine-Tuning Masterclass
**Deep-Dive**: FTDD-02
**Duration**: 45 minutes
**Level**: Senior Engineer and above
**Prerequisites**: FT00 (Steering Stack), FT02 (Open Spectrum), FTDD-01 (MiniCPM)

> *The fully-open comparison. Allen Institute's OLMo and Tülu are what "open" means when a research lab commits to releasing everything — weights, data, code, and the complete post-training recipe. The counterpoint to MiniCPM's product-and-edge orientation.*

---

## Learning Objectives

After this deep-dive, you will be able to:

1. Distinguish Ai2's OLMo (the open base) from Tülu 3 (the open post-training recipe) and explain why both are needed for full reproducibility.
2. Map the OLMo 2 → OLMo 3 progression and the base/instruct/think variants, citing arXiv:2501.00656 for OLMo 2 and arXiv:2411.15124 for Tülu 3.
3. Trace the Tülu 3 post-training pipeline (SFT → DPO → RLVR) and place each stage on the FT00 steering stack.
4. Defend why fully-open (weights + data + code + recipe) matters for reproducibility — and why "weights-only open" is not enough.
5. Contrast the OpenBMB/MiniCPM and Ai2/OLMo-Tülu philosophies: both open-data, but MiniCPM is product/edge-oriented while OLMo/Tülu is research-oriented.

---

## The Subject

The **Allen Institute for AI (Ai2)** is the canonical fully-open lab. Where OpenBMB (FTDD-01) builds small models for product and edge deployment, Ai2 builds models whose *primary contribution is the openness itself* — the demonstration that a frontier-scale model can be released with every component a researcher needs to reproduce, audit, and extend it. Two releases carry this banner: **OLMo** (the open base) and **Tülu 3** (the open post-training recipe).

| Metric | Value |
| --- | --- |
| Lab | Allen Institute for AI (Ai2) |
| License | Apache-2.0 (weights, data, code) |
| Open base | OLMo 2 (arXiv:2501.00656) · OLMo 3 (Nov 2025, 7B + 32B) |
| Open post-train | Tülu 3 (arXiv:2411.15124), up to 405B |
| Pipeline repo | `allenai/open-instruct` |
| OLMo 2 nickname | "2 OLMo 2 Furious" |
| OLMo 3 variants | base · instruct · think |
| Tülu 3 stages | SFT → DPO → RLVR (verifiable rewards) |

Ai2 matters for this course because it is the strict end of the open spectrum (FT02). Where MiniCPM is open-data and open-recipe, OLMo and Tülu push further: they release intermediate checkpoints, the eval harness, and the *exact data mixture*. A skilled person can rebuild the model from scratch; a security team can prove no training-time exfiltration occurred. This is the tier that IL5/IL6 and air-gapped environments increasingly require (FT22).

---

## OLMo — The Open Base

### OLMo 2 — "2 OLMo 2 Furious"

OLMo 2 (arXiv:2501.00656) is the open base that established the modern fully-open standard. The nickname — "2 OLMo 2 Furious" — is the community's nod to the paper's density and the team's pace. What OLMo 2 contributes is not a benchmark win; it is the release of every component the FT02 open-recipe tier promises:

- **Weights** (multiple sizes, base and instruct).
- **Training data** (the actual corpus, not an aggregate description).
- **Training code** (the exact scripts and configs).
- **Intermediate checkpoints** (so you can study training dynamics).
- **Evaluation suite** (so you can reproduce the benchmarks).

The OLMo 2 paper is the canonical "fully open" citation this course returns to in FT02. When the course says "open-recipe means you can rebuild and prove it," OLMo 2 is the example. Every other open-recipe release (SmolLM3, MiniCPM) is measured against the OLMo standard.

### OLMo 3 — the three-variant release

OLMo 3, released November 2025, comes in two sizes (7B and 32B) and — critically — three variants:

- **base** — the pretrained weights, post-training-free. What you would fine-tune from.
- **instruct** — instruction-tuned (SFT + preference). The general-purpose chat model.
- **think** — a reasoning-tuned variant (the GRPO/RL-on-verifiable-rewards lineage, FT14).

The three-variant release is significant because it makes the steering stack (FT00) concrete. The *base* is Layer 1. The *instruct* is base + Layer 3 (SFT + DPO). The *think* is instruct + Layer 3 (GRPO). A student can load all three from the same family and *see* what each steering stage changed — a pedagogical gift that a single-variant release cannot provide. This is why OLMo 3 is the model to reach for when teaching the difference between SFT, DPO, and GRPO: the same base, three steers, directly comparable.

---

## Tülu 3 — The Open Post-Training Recipe

If OLMo is the open base, **Tülu 3** (arXiv:2411.15124) is the open post-training stack. Tülu 3's contribution is the full, reproducible recipe for turning a base model into an aligned, capable instruct model — released as data, code, and configs, at scales up to 405B parameters.

### The three-stage pipeline

Tülu 3 post-training is a three-stage pipeline, and each stage maps cleanly onto the FT00 steering stack:

```
BASE MODEL (Layer 1)
    │
    ▼
[1] SFT  ────────────────── Layer 3 (format, instruction-following)
    │   (Supervised Fine-Tuning on curated instruction data)
    ▼
[2] DPO  ────────────────── Layer 3 (preference alignment)
    │   (Direct Preference Optimization on preference pairs)
    ▼
[3] RLVR ────────────────── Layer 3 (reasoning, verifiable rewards)
    │   (Reinforcement Learning on Verifiable Rewards — math, code, etc.)
    ▼
TÜLU 3 INSTRUCT MODEL
```

**Stage 1 — SFT.** Supervised Fine-Tuning on a curated instruction dataset. This is the format-and-injection-following steer (FT12). Tülu 3 releases the SFT data mixture, so you can see exactly what instructions the model was steered toward.

**Stage 2 — DPO.** Direct Preference Optimization on preference pairs (FT13). This is the preference-alignment steer — "this response is better than that one." Tülu 3 releases the preference dataset and the DPO hyperparameters.

**Stage 3 — RLVR.** Reinforcement Learning on Verifiable Rewards (FT14). This is the reasoning steer — the model generates solutions to problems with verifiable answers (math, code, logic), and correct solutions are rewarded. RLVR is the technique behind the "reasoning model" wave, and Tülu 3 is the open implementation of it.

The load-bearing property: every stage's data, code, and config is released. You can rebuild Tülu 3 from an OLMo base. You can ablate a single stage (what if you skip RLVR?). You can swap in your own data at any stage. This is what "open post-training recipe" means in practice — not a paper describing the recipe, but the recipe itself, executable.

### The open-instruct repo

The practical handle for all of this is `allenai/open-instruct` — the repository that contains the Tülu 3 training code, data preparation scripts, and configuration files. If you want to reproduce Tülu 3, or build a Tülu-style pipeline on your own base, open-instruct is where you start. It is the post-training analogue of OLMo's pretraining code: the executable recipe, not the description.

---

## Why Fully-Open Matters for Reproducibility

The argument for fully-open (weights + data + code + recipe) is not ideological. It is the FT02 argument, sharpened: reproducibility is the property that lets you *prove* things about a model, and reproducibility requires every component.

Three things only fully-open releases give you:

1. **Reproducibility in the strict sense.** You can rebuild the model from a pinned commit and get the same weights (modulo nondeterminism, which is itself documented). This is the no-silent-drift argument (FT02) — you can prove the model you validated is the model you are running.
2. **Ablation.** You can remove or change a single component (a data source, a post-training stage) and measure the effect. This is how research advances — and it is impossible with a weights-only release, where the data and recipe are opaque.
3. **Trust in the supply chain.** You can audit the training pipeline end-to-end and rule out hidden training-time exfiltration. This is the property air-gapped and classified environments require (FT22).

The Tülu 3 release is the clearest demonstration of (2) — ablation — in the open ecosystem. Because the full post-training recipe is released, a researcher can ask "what does RLVR actually contribute?" by training Tülu 3 with and without the RLVR stage, on the same base, with the same SFT and DPO. With a weights-only release, that question is unanswerable.

---

## OpenBMB vs Ai2 — Two Open Philosophies

Both OpenBMB (MiniCPM) and Ai2 (OLMo/Tülu) are open-data and open-recipe. Both ship Apache-2.0. But they have different centers of gravity, and the difference matters when you choose a base.

| Property | OpenBMB / MiniCPM | Ai2 / OLMo + Tülu |
| --- | --- | --- |
| **Orientation** | Product / edge | Research |
| **Model sizes** | Small (1B–4B) + multimodal | Mid-to-large (7B–405B), text-focused |
| **Post-training recipe** | Disclosed, productized | Fully released, ablation-ready (Tülu 3) |
| **Multi-variant releases** | Modality axis (text/vision/omni) | Steer axis (base/instruct/think) |
| **Default use case** | Ship a small model on edge hardware | Reproduce, audit, research the stack |
| **Course role** | The on-ramp base (FT00 lab) | The fully-open reference (FT02, FT22) |

The summary: **MiniCPM is what you ship; OLMo/Tülu is what you study.** If your goal is a small model on a phone or an edge device, MiniCPM's modality axis and small sizes win. If your goal is to understand, audit, or reproduce the full training-and-post-training stack — especially for a regulated deployment that demands proof — OLMo and Tülu are the reference.

This is why the course uses both. FT00's lab loads MiniCPM5-1B because it is cheap to iterate on. FT02's audit references OLMo 2 because it is the fully-open standard. And when the course needs to teach *what post-training actually does* (FT12 SFT, FT13 DPO, FT14 GRPO), OLMo 3's base/instruct/think variants and Tülu 3's released pipeline are the clearest teaching tools — the same base, each steer visible.

---

## Anti-Patterns

### Treating "fully-open" as "production-ready"

OLMo and Tülu are research artifacts first. They are fully open, auditable, and reproducible — but they are not always the most capable models at their size point. A closed or weights-only model may outperform them on benchmarks. Choose them for openness, auditability, and reproducibility, not because they top every leaderboard. Capability is FT03's concern.

### Assuming OLMo 3 "think" = GPT-class reasoning

The think variant is a reasoning-tuned model (RLVR), but it is a 7B/32B model, not a frontier model. It demonstrates the RLVR technique (FT14) openly — which is its research contribution — but it will not match a closed frontier reasoning model on the hardest tasks. Use it to *understand* reasoning fine-tuning, not to deploy a frontier reasoner.

### Skipping the open-instruct repo

Tülu 3's value is in the recipe, and the recipe lives in `allenai/open-instruct`. Reading the paper without reading the repo gives you the description; reading the repo gives you the executable pipeline. If you want to reproduce or adapt Tülu 3, the repo is the primary source.

---

## Key Terms

| Term | Definition |
| --- | --- |
| **Ai2** | Allen Institute for AI; the canonical fully-open lab |
| **OLMo 2** | The open base that established the modern fully-open standard (arXiv:2501.00656); nicknamed "2 OLMo 2 Furious" |
| **OLMo 3** | The three-variant release (Nov 2025, 7B + 32B): base, instruct, think |
| **Tülu 3** | The open post-training recipe (arXiv:2411.15124); SFT → DPO → RLVR, up to 405B |
| **RLVR** | Reinforcement Learning on Verifiable Rewards — the reasoning steer (Tülu 3 stage 3, FT14) |
| **open-instruct** | `allenai/open-instruct` — the repo containing the Tülu 3 training code, data prep, and configs |
| **Fully-open** | Weights + data + code + recipe + checkpoints + eval — the strictest tier (FT02) |
| **base / instruct / think** | OLMo 3's three variants: pretrained, instruction-tuned, reasoning-tuned |

---

## Lab Exercise

See `07-lab-spec.md`. The lab compares OLMo 3's base, instruct, and think variants on a reasoning task — making the three steering stages (none, SFT+DPO, +RLVR) directly visible on the same base, no fine-tuning required.

---

## References

1. **OLMo 2 paper** — arXiv:2501.00656; the fully-open base standard.
2. **Tülu 3 paper** — arXiv:2411.15124; the open post-training recipe.
3. **open-instruct repo** — `allenai/open-instruct`; the executable Tülu 3 pipeline.
4. **OLMo project page** — `allenai.org/olmo`; the family overview.
5. **Course 3 FT02** — the Open Spectrum; OLMo/Tülu as the open-recipe reference.
6. **Course 3 FT14** — GRPO and verifiable rewards; RLVR is the open implementation.
7. **Course 3 FTDD-01** — MiniCPM; the product/edge counterpoint to OLMo/Tülu's research orientation.