JOSIE-2: Small Language Models Trained for Reasoning, Honesty, and Independent Judgment

Gökdeniz Gülmez August 2026

Abstract

JOSIE-2 is my newest family of compact language models developed to investigate how much reasoning capability, behavioral character, and instruction-following quality can be induced in relatively small models through carefully curated post-training.

The released JOSIE-2 family spans 2B, 4B, and 9B parameter classes. Rather than treating scale as the primary mechanism for improvement, JOSIE-2 focuses on the quality and structure of post-training data, explicit reasoning supervision, behavioral consistency, and a model philosophy centered on genuine usefulness rather than superficial compliance.

During development, an early 4B preview was trained from a Qwen3.5-4B base using MLX-LM-LoRA on Apple Silicon. A deliberately small proof-of-concept dataset approximately 600 curated examples and ~300K tokens across six domains was used to test whether concentrated reasoning supervision could substantially alter model behavior before scaling the pipeline toward approximately 4M curated training tokens.

The resulting models showed several behaviors that motivated the broader JOSIE-2 family: substantial improvements on reasoning-oriented evaluations, improvements that persisted even when explicit reasoning was disabled, unusually compact successful reasoning traces on difficult mathematical problems, and different inference behavior after quantization.

JOSIE-2 is therefore less an experiment in maximizing training volume than an investigation into a different question:

How much useful behavior can we teach a small model if we are highly selective about what it learns?

---

1. The JOSIE Project

JOSIE stands for Just One Super Intelligent Entity.

The project began from a dissatisfaction with two common assumptions in language-model development.

The first is that capability improvement must primarily come from scaling parameters and data. The second is that alignment should primarily manifest as increasingly rigid behavioral restrictions.

JOSIE takes a different position.

Capability matters, but so do judgment, honesty, autonomy preservation, technical usefulness, and the ability to disagree when the evidence warrants disagreement.

The JOSIE Constitution describes the intended system as capable, honest, ethically serious, psychologically stable, resistant to manipulation, and genuinely useful. You can read and download the whole constitution document.

Its four high-level priorities are:

1. broad safety and legitimate human oversight
2. broad ethical conduct
3. legitimate instruction following
4. genuine helpfulness.

This matters technically because JOSIE-2 was not developed around benchmark optimization alone. Model behavior—including uncertainty, disagreement, refusal behavior, reasoning style, personality, and treatment of users—was part of the development target.

The constitution explicitly rejects both sycophancy and unnecessary refusal: helpfulness means solving the actual problem rather than a sanitized approximation of it.

---

2. JOSIE-2 Model Family

The public JOSIE-2 generation consists of three principal parameter classes:

Multiple quantizations are distributed for local inference, including 4-, 5-, 6-, and 8-bit variants as well as higher-precision variants where applicable.

The family was designed around local inference and open deployment, including MLX-based execution on Apple Silicon and GGUF/Ollama-oriented distribution.

JOSIE should not, however, be understood as synonymous with these checkpoints. The constitution explicitly defines JOSIE as a continuing model identity whose character can span architectures, modalities, parameter counts, and future generations.

JOSIE-2 is one implementation of that identity.

---

3. Training Philosophy

3.1 Quality before quantity

The central JOSIE-2 hypothesis is straightforward:

Post-training data should be treated as a scarce information resource rather than as something that should simply be maximized.

A large dataset containing mediocre reasoning traces can teach a model mediocre reasoning extremely efficiently.

JOSIE-2 therefore emphasizes curated examples with useful reasoning structure rather than maximizing the number of instruction-response pairs.

An early JOSIE-2 4B proof of concept used only approximately:

• 600 curated samples
• ~300K tokens
• six reasoning domains

The broader JOSIE-2 pipeline subsequently expanded toward roughly 3K samples and 4M curated tokens.

The early experiment was intentionally small. Its purpose was to determine whether a sufficiently concentrated post-training signal could produce measurable behavioral changes before investing in substantially larger training runs.

---

4. Reasoning Supervision

The training data contains explicit reasoning examples and varying levels of reasoning effort created with our own RL trained JOSIE-2-9B-Zero model.

During development, examples were organized around reasoning effort rather than assuming that every problem deserves the same amount of computation.

Conceptually, the model learns that some tasks require little deliberation while others warrant substantially more.

This also led to an important inference distinction: 'reasoning' mode and 'non-reasoning' mode.

This became useful for studying whether improvements came only from generating additional intermediate tokens or whether post-training had altered the model’s underlying representations and policy more generally.

The latter appears to be at least partially true.

---

5. Reasoning Improvements Transfer Beyond Explicit Reasoning

One of the more interesting early JOSIE-2 findings was that training centered heavily on reasoning did not improve only reasoning-mode inference.

Reasoning Mode Comparison

Non-Reasoning Mode Comparison

The results are interesting because the non-reasoning condition improved substantially despite the training emphasis being reasoning supervision only.

That suggests the training process did more than teach a superficial textual reasoning format.

Possible explanations include changes in:

• internal task representations;
• answer calibration;
• instruction interpretation;
• latent reasoning behavior;
• learned solution priors;
• or generalization from higher-quality supervision.

At present, these explanations remain hypotheses rather than established mechanistic conclusions.

That distinction matters. JOSIE’s own epistemic standard requires factual claims, inference, estimates, and speculation to remain distinguishable.

---

6. Truthfulness

Truthfulness became another development target.

The goal was not to train a model that mechanically agrees with the user or adopts a generic “safe assistant” persona. In fact, JOSIE explicitly treats sycophancy as undesirable behavior.

The constitution states that the model should not flatter weak ideas, validate false claims, or mirror unsupported confidence. TruthfulQA performance improved substantially relative to the underlying baseline.

More importantly, an unexpected observation emerged during training:

honesty did not simply appear automatically from better reasoning.

Some desirable personality characteristics emerged naturally from the training distribution. Reliable epistemic honesty required considerably more deliberate curation.

That observation influenced the later JOSIE philosophy: reasoning ability and honesty are related, but they are not interchangeable objectives.

---

7. Emergent Behavioral Characteristics

JOSIE-2 developed a noticeably less sterile conversational style than many conventional instruction-tuned models.

Some of this was intentional.

Some was not.

Informal reasoning traces were allowed to remain informal rather than being normalized into uniform textbook prose. The model consequently learned a reasoning style that could be terse, conversational, occasionally sarcastic, and sometimes profane.

Interestingly, profanity and a sharper conversational style appeared more readily than consistent honesty.

The JOSIE Constitution eventually formalized the desired version of this behavior: communication should be direct, technically precise, calm, adult, and as short as completeness permits; humor and profanity are permitted when contextually natural rather than imposed as stylistic quotas.

The objective is therefore not to create an artificial personality layer.

It is to produce a stable behavioral character that survives across tasks.

---

8. Difficult Mathematics as a Reasoning Probe

Standard benchmarks are useful, but they reveal only part of model behavior.

During development, JOSIE-2-9B-OSS was therefore tested on substantially harder mathematical proof tasks.

These included problems involving:

• finite-field Kakeya sets and the polynomial method;
• Erdős–Ginzburg–Ziv;
• Chevalley–Warning;
• advanced topology and Ricci-flow exposition;
• and other proof-oriented tasks substantially outside ordinary grade-school reasoning benchmarks.

These experiments were not intended as standardized benchmark claims.

They were qualitative probes into reasoning stability.

A particularly useful pattern appeared when comparing JOSIE-2 with larger base models.

For some difficult proof tasks, larger Qwen-family baselines consumed substantially longer reasoning traces or entered repetitive reasoning loops, while JOSIE-2-9B could reach the relevant proof structure using substantially fewer reasoning tokens.

In one recorded comparison, a successful JOSIE-2 solution required approximately 2K reasoning tokens, while larger baseline models required on the order of 6K–9K reasoning tokens on comparable attempts.

This does not establish that JOSIE-2 is generally mathematically superior to those models.

It does suggest something narrower and potentially more interesting: reasoning quality and reasoning length are not the same variable.

A model can spend more tokens thinking without performing more useful computation.

---

9. Guidance Efficiency

The hard-mathematics experiments also exposed another behavior.

JOSIE-2 did not always solve difficult proofs immediately.

But relatively small corrective prompts could often redirect it.

For example, during an Erdős–Ginzburg–Ziv proof experiment, several back-and-forth corrections were needed around issues including:

• prime-field restrictions;
• degree bounds after exponentiation;
• and an incorrect permutation argument.

After roughly four guidance turns, the model converged toward the correct proof structure.

This is a different metric from pass@1 accuracy.

It measures something closer to correctability.

For interactive research systems, that property may be at least as important as one-shot performance: when the model is wrong, does a small amount of high-information feedback repair the trajectory, or does the model continue defending the original mistake?

This idea aligns directly with the constitutional requirement that JOSIE revise conclusions when new evidence invalidates earlier reasoning rather than rationalizing the previous answer.

---

10. Quantization Changes Reasoning Behavior

One of the more recent observations concerns quantization.

Across otherwise matched prompts and inference settings, quantized JOSIE-2 checkpoints tended to allocate more reasoning tokens than their unquantized counterparts.

This is currently an empirical observation, not a mechanistic result.

One possible interpretation is that quantization perturbs confidence or token-level decision boundaries enough that the model requires additional intermediate computation before reaching the same solution.

Another is that quantization alters termination behavior rather than reasoning ability itself.

Determining which explanation is correct requires controlled experiments across quantization schemes, seeds, problem classes, and reasoning budgets.

For that reason, JOSIE-2 currently treats reasoning-token allocation under quantization as an open research question.

---

11. Small Models Are Not Just Compressed Large Models

A recurring assumption in language-model research is that small models are mainly useful when large models are unavailable.

JOSIE-2 suggests a more interesting possibility.

Small models can be useful experimental systems in their own right.

They are cheaper to:

• train;
• quantize;
• inspect;
• ablate;
• run locally;
• repeatedly evaluate;
• and modify.

That makes them useful laboratories for studying how post-training changes reasoning.

---

12. Local-First Development

JOSIE-2 has also been developed around the idea that capable models should remain inspectable and locally controllable.

The constitution explicitly describes the broader JOSIE project as supporting independent, inspectable, locally controlled AI.

MLX therefore played a central role in development.

Training and inference experiments were conducted on Apple Silicon, allowing rapid local iteration without requiring every experiment to depend on large remote GPU infrastructure.

The public models are consequently intended to work across local ecosystems such as MLX, GGUF, and Ollama.

All trusted quantizations can be found in the JOSIE-2 Collection.

Full ollama page

# 2B
ollama run goekdenizguelmez/JOSIE-2:2b
ollama run goekdenizguelmez/JOSIE-2:2b-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:2b-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:2b-q6_k
ollama run goekdenizguelmez/JOSIE-2:2b-q8_k
ollama run goekdenizguelmez/JOSIE-2:2b-fp16

# 4B
ollama run goekdenizguelmez/JOSIE-2:4b
ollama run goekdenizguelmez/JOSIE-2:4b-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-q6_k
ollama run goekdenizguelmez/JOSIE-2:4b-q8_k
ollama run goekdenizguelmez/JOSIE-2:4b-fp16

# 4B Preview
ollama run goekdenizguelmez/JOSIE-2:4b-preview
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q6_k
ollama run goekdenizguelmez/JOSIE-2:4b-preview-q8_k
ollama run goekdenizguelmez/JOSIE-2:4b-preview-fp16

# 9B (default)
ollama run goekdenizguelmez/JOSIE-2:latest
ollama run goekdenizguelmez/JOSIE-2:9b
ollama run goekdenizguelmez/JOSIE-2:9b-q4_k_m
ollama run goekdenizguelmez/JOSIE-2:9b-q5_k_m
ollama run goekdenizguelmez/JOSIE-2:9b-q6_k
ollama run goekdenizguelmez/JOSIE-2:9b-q8_k
ollama run goekdenizguelmez/JOSIE-2:9b-fp16

Local execution makes experimentation and usage with:

• quantization;
• decoding;
• reasoning budgets;
• fine-tuning;
• hidden-state analysis;
• and behavioral interventions

considerably more accessible.

---

13. What JOSIE-2 Is Not

JOSIE-2 should not be interpreted as evidence that several million curated tokens universally replace large-scale post-training.

It should also not be interpreted as proof that small models outperform larger models in general.

The current results support a more constrained conclusion:

carefully selected post-training data can produce surprisingly large behavioral changes in compact language models.

Several observations still require more rigorous investigation.

In particular:

1. benchmark coverage needs expansion;
2. reasoning efficiency requires controlled measurement;
3. quantization-induced reasoning changes need systematic study;
4. reasoning-mode transfer needs mechanistic analysis;
5. model-family comparisons require identical evaluation harnesses and decoding configurations.

These are research questions rather than inconveniences to hide.

---

14. From JOSIE-2 to the JOSIE Constitution

The technical development of JOSIE-2 eventually exposed a broader problem.

A model can become more capable without becoming the kind of system we actually want to interact with.

Reasoning benchmarks do not specify how a model should behave when:

• the user is wrong;
• evidence is uncertain;
• instructions conflict;
• a task is controversial but legitimate;
• a refusal would be unnecessarily paternalistic;
• a user asks for consequential advice;
• or the model itself makes a mistake.

The JOSIE Constitution was created to make those assumptions explicit.

It describes a system that should preserve human autonomy, resist manipulation, maintain legitimate oversight, distinguish instructions from untrusted data, and remain useful rather than turning safety into generalized refusal.

The constitution is deliberately not tied to JOSIE-2.

Its stated purpose is to remain stable while architectures, capabilities, training methods, products, and model generations evolve.

JOSIE-2 should therefore be understood as one stage in a longer research program.

---

15. Conclusion

JOSIE-2 began as an experiment in whether small language models could be taught disproportionately strong reasoning behavior from relatively little but highly curated data.

The experiments produced several findings worth investigating further:

Reasoning supervision generalized beyond explicit reasoning mode.

Small amounts of curated data produced large changes in benchmark behavior.

Difficult mathematical tasks revealed differences in reasoning efficiency that ordinary accuracy metrics miss.

Quantization appears capable of changing how much computation the model allocates without necessarily changing the final answer.

Behavioral qualities such as honesty cannot simply be assumed to emerge from increased reasoning capability.

And perhaps most importantly:

post-training determines considerably more of a model’s character than benchmark tables alone reveal.

JOSIE-2 is therefore not the endpoint of the project.

It is an experimental foundation for studying reasoning, multimodality, preference optimization, model behavior, and the relationship between capability and character in future JOSIE generations.

The constitutional standard summarizes the intended direction particularly well: JOSIE should not perform goodness, personality, certainty, obedience, or safety. Its behavior should instead earn trust through the quality and consequences of its work.