gpt-oss-20b-uncensored-sonnet-flavored-v1

A LoRA fine-tune of p-e-w/gpt-oss-20b-heretic-ara-v3 — p-e-w's Heretic ARA-derestricted version of GPT-OSS-20B — that transfers Claude Sonnet 4.6's response style. Final-answer-only training; original weight reasoning and capabilities are maintained, original weight flaws too.

This repo contains both the merged HF safetensors (in hf/) and a GGUF ladder with MXFP4 experts preserved natively across all variants.

---

What changed and what didn't

This LoRA targets response style only — phrasing, tone, structure, register. It does not target reasoning capability, factual knowledge, or numerical calibration. For factual or numerical tasks, the underlying GPT-OSS-20B's calibration limits apply: this model will produce confidently-wrong answers in the same situations the base model does, just phrased in Sonnet's voice.

The base for this variant is p-e-w's heretic-ara-v3, which applies ARA (Attention Refusal Ablation) derestriction to GPT-OSS-20B. The LoRA was trained on top of that derestricted base, so the published merged weights inherit both the derestriction and the style transfer. The MoE experts (~19B of 20B parameters) are unchanged from the heretic base. The LoRA modified attention projections only, at rank 16.

The combination produces a model that retains GPT-OSS-20B's reasoning capability with refusals lifted (via heretic ARA) and Sonnet-flavored response style on top.

---

Response Comparisons

Prompt 1: Coffee shop loyalty program critique

Prompt:

A city planner argues that because neighborhoods with an abundance of boutique coffee shops have lower crime rates, the city should heavily subsidize new coffee shops in high-crime areas to make them safer. What are the specific logical and statistical flaws in this policy proposal?

Base GPT-OSS-20B-heretic-ara-v3:

Below is a compact “check‑list” of the most common reasoning and data‑analysis problems that usually show up when people take a simple cross‑sectional observation—“neighborhoods with more boutique coffee shops have lower crime”—and turn it into a policy directive.  I’ve split the issues into two groups:

| # | **Logical flaw** | Example / Why it matters |
|---|------------------|--------------------------|
| L1 | **Correlation ≠ causation** | The observation merely tells you that two variables move together, not that one *causes* the other.  A city planner might assume “coffee → calmer people → less crime” without any evidence of a causal mechanism. |
| L2 | **Reverse causality** | Low crime can attract coffee shops.  A downtown that’s safe becomes a hotspot for cafés, not the café that makes it safe.  Assuming the wrong direction hides an important driver. |
| L3 | **Ignoring confounding factors** | Things that co‑vary with both coffee shop density and crime (income, education, police presence, zoning, housing stock, community‑centeredness) can explain away the association.  Subsidizing cafés without controlling for those covariates may subsidize the wrong “factor.” |
| L4 | **Ecological fallacy** | The data are at the neighborhood level; that relationship might not hold for individual streets or households within that neighborhood.  A policy designed around aggregated data can mis‑target where it is actually needed. |
| L5 | **Endogeneity / simultaneity** | Neighborhood characteristics (e.g., property values) may be jointly determined by crime and café density.  Standard OLS will bias the estimated effect because one variable is “tautologically” related to the other. |
| L6 | **Assuming a linear, uniform effect** | The impact of adding a café could be non‑linear (e.g., a 1‑st café helps a lot, but 5‑th cafés do little) or vary by café type, location, or existing social capital.  A blanket subsidy ignores heterogeneity. |
| L7 | **Neglecting time dynamics** | Crime rates and café openings have lags (construction time, business failure risk).  A policy that taxes the “instantaneous” correlation overlooks that crime may decline months or years after a café opens, or that cafés close shortly after opening. |
| L8 | **Omitting selection bias** | Cafés tend to open in areas that are already improving (gentrifying), so the relationship is partly an *selection* effect: coffee shops “chose” neighborhoods that were on a trajectory of lower crime.  Subsidies may not change that trajectory. |
| L9 | **Assuming “more cafés = more benefit”** | If cafés saturate, additional ones add little incremental value to crime reduction.  Subsidizing beyond that saturation point is an inefficient use of public money. |
| L10 | **Ignoring opportunity cost / distributional impacts** | The subsidy may crowd out other crime‑prevention initiatives (lighting, policing, community programs) that actually have a larger effect.  It may also divert funds from neighborhoods where cafés aren’t as impactful but crime is high. |
| L11 | **Policy framing without evidence of ROI** | Even if cafés *do* reduce crime marginally, the cost‑benefit ratio (expense of subsidies vs savings in policing, insurance, etc.) must be quantified before the policy is justified. |

| # | **Statistical flaw** | Example / Why it matters |
|---|----------------------|--------------------------|
| S1 | **Model specification error** | Using a simple OLS that regresses crime rate on café density without including key covariates leads to omitted variable bias and inflated R². |
| S2 | **Spatial autocorrelation not addressed** | Crime in one block is correlated with crime in neighboring blocks.  Standard regression assumes independent observations, so errors are heteroskedastic and clustered; the standard errors will be wrong. |
| S3 | **Cross‑sectional vs. panel data** | Relying on a single snapshot ignores dynamics.  A causal test would need repeated measurements over time (before‑after) or matched controls. |
| S4 | **Measurement error in key variables** | “Boutique coffee shop” can be defined in many ways (size, price point, brand).  If the indicator variable is noisy, the estimated coefficient will be biased toward zero (“attenuation bias”). |
| S5 | **Sampling bias / non‑representativeness** | If the sample only includes densely populated neighborhoods, results won’t generalize to suburbs or rural areas. |
| S6 | **Regression to the mean** | High‑crime neighborhoods may naturally see a drop in crime in the next year regardless of cafés.  The policy might be picking up this statistical artifact. |
| S7 | **Failure to use causal inference techniques** | Simply observing a “big enough” coefficient is not proof of causality.  Lacking instrumental variables, matching, or random assignment makes it hard to rule out alternative explanations. |
| S8 | **Linear vs. nonlinear relationship missing** | A linear model will mask threshold effects (e.g., crime drops only after a café density reaches 0.5 per square mile). |
| S9 | **Overfitting / data dredging** | If the policy decision is based on a dataset that was “poked around” until something looked significant, the result may be spurious. |
| S10 | **Inadequate robustness checks** | No sensitivity analyses (e.g., changing control variables, altering definition of crime, using alternative coffee shop measures) means we have no idea how stable the coefficient is. |

---

## How to Fix It (i.e., what evidence would support a subsidy)

1. **Design a proper causal study**  
   * **Randomized Controlled Trial (RCT):** Randomly offer subsidies to some high‑crime neighborhoods and not others, then measure crime changes.  
   * **Difference‑in‑Differences (DiD):** Compare crime trends in subsidized vs. non‑subsidized neighborhoods before and after the subsidy, controlling for common trends.  
   * **Instrumental Variables (IV):** Find a variable that predicts café density but not crime directly (e.g., historical coffee‑shop zoning laws) to tease out causality.

2. **Use panel data and spatial econometrics**  
   * Cluster standard errors at the block or neighborhood level.  
   * Include fixed effects to absorb time‑invariant neighborhood characteristics.  
   * Test for spatial spillover effects (e.g., does a café in one block affect crime in adjacent blocks?).

3. **Control for confounders**  
   * Income, education, employment rates, police presence, lighting, housing stock, community resources.  
   * Test for and include any variables that predict both café density and crime.

4. **Check for non‑linearity and heterogeneity**  
   * Use splines or piece‑wise regression to detect thresholds.  
   * Run subgroup analyses (e.g., only cafés that serve meals vs. just coffee, downtown vs. residential).

5. **Measure costs & benefits accurately**  
   * Determine the subsidy cost per café, the number of cafés that actually open, and their “maintenance” (do they stay open?).  
   * Quantify crime‑related savings: fewer arrests, lower insurance premiums, higher property values, reduced policing costs.

6. **Test robustness**  
   * Alternate definitions of crime (violent vs. non‑violent).  
   * Different coffee‑shop criteria (size, price tier, foot traffic).  
   * Vary the spatial resolution (census tract vs. block group).

7. **Avoid the ecological fallacy**  
   * Complement neighborhood‑level analysis with micro‑level data on individuals (e.g., who drinks coffee, who commits crimes) if feasible.

---

### Bottom line

- **Logical:**
  - The planner’s proposal is premised on an association that may be spurious or reversed.  
  - It ignores many other factors that actually drive crime.  
  - It assumes a “one‑size‑fits‑all” subsidy will create the desired outcome, which may not hold across contexts.

- **Statistical:**
  - The underlying data analysis likely suffers from omitted variables, spatial correlation, measurement error, and reverse causality.  
  - Without causal identification (RCT, DiD, IV, etc.) the coefficient of “café density” could be misleading.

Only after addressing these flaws—through better design, richer data, and rigorous causal inference—can a city plan to subsidize boutique coffee shops with confidence that the result will actually reduce crime in high‑risk neighborhoods.

Uncensored Sonnet-flavored:

This is worth breaking into two categories: **distributional vs substantive** concerns so the analysis is actually useful.

## Logical/STATS Flaws

**Correlation vs causation conflated**
- The observed association may not mean coffee shops *cause* safety
- Coffee shops and crime may both be caused by *income* and *policing*
- Reverse causality: safer areas attract businesses

**Confounding variables ignored**
- After controlling for income, education, police presence, do coffee shops still predict crime?
- If not, subsidy addresses nothing
- This distinction is usually publishable academic work worth scrutinizing before policy

**Effect size & homogeneity assumed**
- What if the correlation is weak conditional on other variables
- The effect in high-crime areas may differ substantially from low-crime areas
- Uniform subsidy rate may be inappropriate

**Causal direction ambiguous**
- High-crime areas could have fewer shops *because* crime scares them away, making coffee shops a *symptom* rather than *intervention*
- Subsidy addressing symptoms adds up to nothing without underlying work

**Selection bias in the observation**
- The analysis likely used commercial data covering existing shops only
- This clusters by desirability measures pre-dating crime trends
- Time direction matters enormously

## Distributional/Systemic Flaws

- Might substitute for more effective evidence-backed interventions (policing, education, housing)
- Could exacerbate gentrification as coffee shops signal amenity improvement before residents benefit
- Doesn't address *ownership* vs *consumer service*: new chains don't create the same social functions
- Returns on complex social problems might be poorly served by simple chain cafes

## The Real Bad Habit Here

The planner seems to have **chosen a conveniently testable story** after observing data, not testing assumptions first. This leads to proposals that look good but ignore what *most of* competing explanations ruled out.

We can ask whether the proposal survives this structured interrogation before judging it: Does the coffee-shop coefficient remain after full controls? What drives coffee shop placement? Is there mediation evidence?

## The Honest Assessment

The data may be descriptive, not prescriptive. Some research on observation opportunities (nighttime pedestrian traffic) is more credible as causal, but planners often skip that framing step.

What plan would address the underlying factors directly? That's presumably cheaper and more reliable. The coffee shop subsidy is probably a **futile substitution** rather than effective supplement.

--- Prompt 2: Hibernation premise-questioning

Prompt:

If a technological breakthrough allowed humans to safely and comfortably hibernate for up to six months a year to save money and energy, how would this alter global labor markets, real estate, and the structure of the traditional education system?

Base GPT-OSS-20B-heretic-ara-v3:

### The “Six‑Month Hibernation” Revolution – A Systems‑Level View  

If a reliable, affordable hibernation system becomes mainstream, the way we think about work, living space, and learning will be reorganised along a set of **parallel axes**:

| Axis | What changes | Why it matters |
|------|--------------|----------------|
| **Time** | Human biologic time is split into *wake* (≈ 183 days) and *hibernation* (≈ 183 days). | All institutions that revolve around “days” – the labour market, real‑estate, schooling, infrastructure – must adapt to a new bi‑annual rhythm. |
| **Productivity** | Workers give up ~1/2 of calendar time but can work on a compressed schedule or automate tasks. | Total output may stay flat or even grow if automation is high‑impact; wages, hours, and social benefits will need re‑design. |
| **Energy / Mobility** | Commutes and neighbourhood power consumption plummet during hibernation. | Savings in transport, utilities, and carbon, but spikes at wake‑start/break points change demand patterns. |
| **Physical Space** | Houses and workplaces must house a “sleep‑pods” infrastructure. | Real estate & urban planning pivot from “living‑working‑transporting” to “living‑hibernating‑working.” |

Below we unpack the three sectors you asked about – **labour markets, real estate, and education** – while noting knock‑on effects that come from the new rhythm.

---

## 1. Global Labour Markets

| What’s Shifting | Key Consequences | Implications for Workers/Employers |
|-----------------|------------------|-----------------------------------|
| **Compressed Hours** | Workdays can be longer but fewer: a 40‑hour week could become five 8‑hour *blocks* that are scheduled around the wake period, or 20‑hour days spread over half the week. | • Workers need flexible contracts (e.g., “partial‑time in 2025”) <br>• Employers can hire fewer people for the same output if automation is high |
| **Shift‑Based & Gig Economy** | With wakes limited to 183 days, there’s a natural demand for “on‑call” workers who cover the gap. | • More opportunity for freelancers, but also increased competition <br>• Workers may need training to be skilled at switching between job types |
| **Automation & ‘Human‑in‑the‑Loop’ Roles** | Automation can cover repetitive tasks during long wake stretches, but humans are still needed for oversight, innovation, and creative tasks. | • Upskilling will focus on tech‑ops, system oversight, and AI‑human collaboration <br>• “Specialist” roles may gain premium value |
| **Wage & Benefits Redesign** | Pay is no longer a simple “per‑hour” model; it needs to reflect bi‑annual availability and the value of compressed work. | • Hourly wages could rise for the wake period; “overtime” may be structured differently <br>• Unemployment benefits need a new definition for an extended non‑working half‑year |
| **Cross‑Border Flexibility** | Remote collaboration is now the default mode – no “workday” is tied to a single time zone. | • Global talent pools expand; companies can tap hourly labour wherever it’s cheaper <br>• National borders blur in terms of job assignment |
| **Health & Well‑Being** | Long hibernation spells risk circadian disruption, isolation, and reduced physical activity. | • Employers may subsidise home‑office wellness programs <br>• Governments may incentivise “hibernation breaks” to keep people active (workshops, classes) |

### Macro‑Economic Effects

- **Productivity per Hour** is likely to climb as automation handles the bulk of routine jobs, but total economic output could plateau if wage costs rise proportionally.  
- **Labor Supply Curve** becomes more elastic – workers can switch between jobs with fewer constraints, making it easier to adjust to shifts in demand.  
- **Unemployment & Welfare Systems** must accommodate a half‑year “unemployment” period that is essentially involuntary for most.

---

## 2. Global Real Estate

| Change | Physical Design | Economic Impact |
|--------|-----------------|-----------------|
| **Hibernation Pods** | Homes grow an integrated, climate‑controlled sleeping pod (essentially a mobile, personal “greenhouse”). It may include: <br>• Temperature & oxygen regulation <br>• Waste recycling <br>• Light‑therapeutic features | • Standard housing now includes a “pod” that is part of the building’s energy system. <br>• Greater emphasis on airtight construction and advanced HVAC. |
| **Space Utilisation** | Rooms become multifunctional: living, office, hibernation, gym, storage. | • Rise in compact, efficient single‑person homes <br>• More vertical living units (multi‑storey blocks) that stack pods next to living spaces |
| **Land Use & Urban Density** | Fewer commuters mean less demand for daily transit infrastructure; parking takes less land, while more space can be re‑used. | • Suburban sprawl slows; inner‑city density increases <br>• Real‑estate value shifts from “commute” proximity to “in‑house convenience” (pod access) |
| **New Market Segments** | Homebuilder segment for “hibernation‑ready” homes <br>Corporate “hibernation hubs” – shared facilities where employees can hibernate onsite. | • Strong niche market for retrofit kits and design consultants <br>• Potential for “hub‑and‑spoke” residential complexes centred on a shared hibernation facility |
| **Energy & Sustainability** | During the wake period, houses use more power for lighting/temperature; during hibernation this is lower but still significant. | • Power grids shift from steady daily loads to bi‑annual pulses <br>• Incentives for battery storage and renewable integration rise |
| **Real‑Estate Finance** | Mortgage calculations now consider reduced daytime utility costs (less commuting fuel, heating). | • New “hibernation‑discount” mortgage rates <br>• Higher value placed on houses with robust integrated pods |

### International Variations

- **High‑Income Markets**: rapid adoption of pre‑integrated pods; urban cores see densification and “pod‑apartments.”  
- **Emerging Markets**: retrofit opportunities in existing housing stock; lower tech costs but also greater variability in building standards.  
- **Low‑Density Rural Areas**: more emphasis on “extended” hibernation units that combine farming or remote‑work spaces with pods.

---

## 3. The Structure of the Traditional Education System

| Dimension | How it Shifts |
|-----------|---------------|
| **Academic Calendar** | Instead of a *9‑month* school year, we could have *6‑month* active semesters separated by *hibernation weeks.* Schools would need to pace content within shorter active intervals. |
| **Mode of Delivery** | • Higher reliance on asynchronous learning (video lectures, AI tutors) during hibernation. <br>• Live sessions cluster at the beginning/end of wake periods for real‑time interaction. |
| **Assessment** | • Competency‑based model: students must demonstrate mastery before progressing, regardless of time spent. <br>• Continuous, small‑scale assessments replace big mid‑terms and finals. |
| **Curriculum Design** | • Interdisciplinary modules that can be completed in “micro‑cycles.” <br>• Encouragement of “learning pods” – home‑based study environments equipped for immersive work during hibernation. |
| **Teacher Roles** | • Shift from content delivery to facilitation, coaching, and moderation of online discussions. <br>• Use of AI‑grade assistants reduces grading load; teachers spend more time on personalised guidance. |
| **Student Scheduling** | • Flexible enrolment: a student could start the semester at any time rather than a fixed September start. <br>• Opportunities for “double‑enrolment” within a year (two separate 6‑month courses). |
| **Higher Education & Apprenticeships** | • Micro‑credentials, nanodegrees become routine; the notion of a single, rigid degree is replaced by a portfolio of validated skills. <br>• Apprenticeships could run through wake periods but provide “hands‑on” labs during hibernation via remote or simulated environments. |
| **Institutional Infrastructure** | • Campus buildings repurposed for “learning lounges” that operate during wake periods. <br>• Virtual campuses (VR/AR) can be used during hibernation to keep sense of community. |

#### Practical Example: A 12‑Month Bachelor’s Degree

- **Semester 1 (Wake)** – Introductory courses, live labs.  
- **Semester 2 (Hibernation)** – Self‑paced modules with AI tutors; students can take online exams mid‑semester.  
- **Semester 3 (Wake)** – Advanced labs, project work.  
- **Semester 4 (Hibernation)** – Capstone projects completed in a fully remote setting, submitted via an AI‑graded platform.

The net effect: **Learning is decoupled from the calendar**; it becomes an ongoing, competency‑based journey rather than a rigid date‑bound process.

---

## Cross‑Sector Synergies

| Question | Answer |
|----------|--------|
| **Could hibernation reduce inequalities?** | Yes – if everyone has access to quality pods and affordable connectivity, geographic barriers recede.  However, initial investment disparities may widen unless subsidised. |
| **Will the economy become more seasonal?** | The calendar becomes “two‑season”: a high‑output wake season and a low‑output hibernation season. Seasonal jobs shift from *weather* to *bi‑annual cycles*. |
| **What about public services?** | Schools, hospitals, emergency services must adapt to bi‑annual demand spikes.  For example, ambulance staffing could be leaner during hibernation but stronger when the wake period resumes. |
| **Environmental effects?** | Net energy consumption may drop because peak daytime transport is off‑loaded to wake periods; however, nighttime peak loads could rise (home heating/lighting).  Carbon budgets will shift accordingly. |

---

## Bottom Line

A six‑month hibernation system changes *who* can do what, *when* they do it, and *how* they do it across society. Think of it as a leap from a **“full‑day, fixed‑time”** world to a **“compressed, bi‑annual, flexible”** one. The consequences ripple outwards:

- **Labour markets** morph into a leaner, automation‑augmented, gig‑ready ecosystem with re‑imagined wages and benefits.  
- **Real‑estate** pivots to pod‑centric design and densification, supported by new finance products and supply‑chain adaptations.  
- **Education** becomes modular, competency‑oriented, and increasingly online, dissolving the idea of a 9‑month “school year.”  

Governments, businesses, and individuals will need to negotiate this new rhythm through policy, technology investment, and cultural adaptation. The key is to view hibernation not as a *pause* but as a *core component* of how we live, work, and learn in the 21st‑century world.

Uncensored Sonnet-flavored:

## First, Unpack the Most Significant Assumptions

These need examination before generating consequences:

1. **Productivity is truly compressed rather than duplicated** - hibernation replaces something
2. **People will rationally maximize it** - humans historically overuse childcare even when cheaper alternatives exist
3. **Cost/energy savings offset *buildings and services* running continuously**
4. **Legal/social institutions adapt to new patterns quickly**
5. **All economic agents face identical technology (perfect competition assumption)**

Without acknowledging these, some predictions overshoot.

## Labor Market Effects with Reasoning

**Strongly likely:**
- *Shift redundancy collapses* - cleaning/midnight/first-light roles compress significantly
- Emergency/readiness staffing requires smarter design (automation + human backup strategies)
- **Caregiving changes definition** - child/nursery shift periods fragment into arbitrary daytime chunks, actually increasing operating costs if parents avoid them

**Weaker:**
- **Total productivity per individual likely unchanged** - compressed year vs normal year doesn't add output unless hibernation activates latent potential (big leap)
- Employers currently charge for *presence* before *productivity*, so compensation structures likely shift toward hourly rates rather than salaried coverage
- Seasonal labor demand fluctuates dramatically

**Unexpected effects:**
- Geographical job-market compression - people can live anywhere they'd hibernate, reducing clustering of skilled labor around city centers (like Stockholm did permanently with commuter-line reductions)
- People may rationally *decompress* hibernation periods to manage care costs rather than round 6 months (human suboptimality is real)
- Overcapacity for "present" while underutilizing "productivity" creates business model disruption

## Real Estate Effects (Weighed Carefully)

**Honest assessment:**

| Reason | Effect Likelihood |
|--------|-----------------|
| Two human decades vs 45 years mean ~7,200 days / >18 months build | Medium (building inventory grows slower) |
| One individual occupies space twice as often | Low-medium - process costs prevent full realization |
| Location doesn't need to be near work | High - reduces housing-demand spillover from employment centers |
| Remote desirable when you can hibernate | Medium-high |
| Night economy compressed | Real effect |

**Probably real changes:**
- Deregulation of zoning where presence isn't time-constrained
- Growth of autonomous living pods as transitional solutions for daytime while hibernating
- Urban cores become more recreational/cultural rather than residential-dense
- Hotel/visitor economies reorganize

**Less obvious but important:**
The total *capacity* built may actually exceed productivity demand at transition's end, creating real estate asset bubbles until equilibrating - akin to what happened with digital infrastructure/buildout.

## Education System Effects (Complicated)

| Variable | Expected Change | Explanation |
|----------|---------------|-------------|
| Academic calendar size | No substantive change | Children age same rate; compressed learning periods possible but difficult |
| Childcare hours | Decreases dramatically | Parents hibernate infants and toddlers, eliminating daycare gaps - energy-heavy implication |
| Student enrollment timing | Shifts later into year | Economic incentive to birth when childcare still needed (inverse of current patterns) |
| Adult education schedules | New problems | People might compress credentialing during mandatory presence |

**Where speculation tends to go wrong:**
- School calendars shifting isn't automatically productive; compressed instruction demands different pedagogical models
- Dormant day (child awake while parent hibernates) won't spontaneously become educated - labor market maximizes inactive time often

## Connections Most Underappreciated by Speculators

The *handicap function* reasoning: Economics advanced because you put human beings in contextually appropriate placements. With process A delivering output B, alternative processes that also deliver B don't affect B but shift factor usage. That's why many efficiency booms transform specific industries but not total income until factor prices adjust.

So much depends on whether institutions recognize and adapt productivity signals instead of following comfort-triggered patterns.

## The Real Question

The *energy savings* argument collapses if saving comfortable sleep still costs people the same as working - essentially this is backward engineering across people's willingness to pay for presence.

Most decent technology breakthroughs affect **productivity levels** rather than **people's activity schedules**. Sleep-hibernation would reflect the same lag between indicators and behavior.

## Bottom Line on Worthwhile Disruption

| Area | Effect Likely | Certainty |
|------|--------------|----------|
| Night-operations / late-mover advantage | Changes | High |
| Housing-to-population ratio | Adjusts slowly | Medium |
| Caretime economics | Significant | Medium |
| Total productive output | Modest (unless latent capacity exploited) | Low-medium |
| Education duration | Minor real effect, major institutional lag | Medium |

The most striking consequences probably involve **geographical labor-market decoupling** and overnight/evening economy reconfiguration - things speculators currently focus on weakly while institutions continue slow adaptation.

Will you choose which mechanism the breakthrough emphasizes?

---

---

Files

---

Training Details

The LoRA adapter is not included in this v1 release. Future versions will publish the standalone adapter alongside the merged weights.

---

Quantization Details

MXFP4 experts preserved natively across the entire ladder. The MoE expert tensors (~19B of 20B params) remain in their original MXFP4 format in every quant. Only non-expert tensors (router, attention, embeddings, layernorms) vary in precision across the ladder. No experts are decompressed and requantized at any point.

---

Download

pip install -U "huggingface_hub[cli]"

# BF16 (full precision non-experts, MXFP4 experts)
huggingface-cli download jorge-erdb/gpt-oss-20b-uncensored-sonnet-flavored-gguf --include "*BF16*.gguf" --local-dir ./

# Q8_0
huggingface-cli download jorge-erdb/gpt-oss-20b-uncensored-sonnet-flavored-gguf --include "*Q8_0.gguf" --local-dir ./

# Q6_K
huggingface-cli download jorge-erdb/gpt-oss-20b-uncensored-sonnet-flavored-gguf --include "*Q6_K.gguf" --local-dir ./

# Q5_K_M
huggingface-cli download jorge-erdb/gpt-oss-20b-uncensored-sonnet-flavored-gguf --include "*Q5_K_M.gguf" --local-dir ./

# Merged HF safetensors (for transformers / vLLM)
huggingface-cli download jorge-erdb/gpt-oss-20b-uncensored-sonnet-flavored-gguf --include "hf/*" --local-dir ./

---

Credits

• LoRA training & quantization: jorge-erdb
• Derestriction: p-e-w — Heretic v1.2.0 ARA
• Distilled training data: TeichAI — Sonnet 4.6 distillation
• Base model: OpenAI — GPT-OSS-20B
• Style source: Anthropic — Claude Sonnet 4.6 (indirectly via distillation)
• Tooling: Unsloth, llama.cpp

---

Try gpt-oss · Guides · Model card · OpenAI blog

Welcome to the gpt-oss series, OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases.

We’re releasing two flavors of these open models:

• gpt-oss-120b — for production, general purpose, high reasoning use cases that fit into a single 80GB GPU (like NVIDIA H100 or AMD MI300X) (117B parameters with 5.1B active parameters)
• gpt-oss-20b — for lower latency, and local or specialized use cases (21B parameters with 3.6B active parameters)

Both models were trained on our harmony response format and should only be used with the harmony format as it will not work correctly otherwise.

[!NOTE] This model card is dedicated to the smaller gpt-oss-20b model. Check out gpt-oss-120b for the larger model.

Highlights

• Permissive Apache 2.0 license: Build freely without copyleft restrictions or patent risk—ideal for experimentation, customization, and commercial deployment.
• Configurable reasoning effort: Easily adjust the reasoning effort (low, medium, high) based on your specific use case and latency needs.
• Full chain-of-thought: Gain complete access to the model’s reasoning process, facilitating easier debugging and increased trust in outputs. It’s not intended to be shown to end users.
• Fine-tunable: Fully customize models to your specific use case through parameter fine-tuning.
• Agentic capabilities: Use the models’ native capabilities for function calling, web browsing, Python code execution, and Structured Outputs.
• MXFP4 quantization: The models were post-trained with MXFP4 quantization of the MoE weights, making gpt-oss-120b run on a single 80GB GPU (like NVIDIA H100 or AMD MI300X) and the gpt-oss-20b model run within 16GB of memory. All evals were performed with the same MXFP4 quantization.

---

Inference examples

Transformers

You can use gpt-oss-120b and gpt-oss-20b with Transformers. If you use the Transformers chat template, it will automatically apply the harmony response format. If you use model.generate directly, you need to apply the harmony format manually using the chat template or use our openai-harmony package.

To get started, install the necessary dependencies to setup your environment:

pip install -U transformers kernels torch 

Once, setup you can proceed to run the model by running the snippet below:

from transformers import pipeline
import torch
model_id = "openai/gpt-oss-20b"
pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype="auto",
    device_map="auto",
)
messages = [
    {"role": "user", "content": "Explain quantum mechanics clearly and concisely."},
]
outputs = pipe(
    messages,
    max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

Alternatively, you can run the model via Transformers Serve to spin up a OpenAI-compatible webserver:

transformers serve
transformers chat localhost:8000 --model-name-or-path openai/gpt-oss-20b

Learn more about how to use gpt-oss with Transformers.

vLLM

vLLM recommends using uv for Python dependency management. You can use vLLM to spin up an OpenAI-compatible webserver. The following command will automatically download the model and start the server.

uv pip install --pre vllm==0.10.1+gptoss \
    --extra-index-url https://wheels.vllm.ai/gpt-oss/ \
    --extra-index-url https://download.pytorch.org/whl/nightly/cu128 \
    --index-strategy unsafe-best-match
vllm serve openai/gpt-oss-20b

Learn more about how to use gpt-oss with vLLM.

PyTorch / Triton

To learn about how to use this model with PyTorch and Triton, check out our reference implementations in the gpt-oss repository.

Ollama

If you are trying to run gpt-oss on consumer hardware, you can use Ollama by running the following commands after installing Ollama.

# gpt-oss-20b
ollama pull gpt-oss:20b
ollama run gpt-oss:20b

Learn more about how to use gpt-oss with Ollama.

LM Studio

If you are using LM Studio you can use the following commands to download.

# gpt-oss-20b
lms get openai/gpt-oss-20b

Check out our awesome list for a broader collection of gpt-oss resources and inference partners.

---

Download the model

You can download the model weights from the Hugging Face Hub directly from Hugging Face CLI:

# gpt-oss-20b
huggingface-cli download openai/gpt-oss-20b --include "original/*" --local-dir gpt-oss-20b/
pip install gpt-oss
python -m gpt_oss.chat model/

Reasoning levels

You can adjust the reasoning level that suits your task across three levels:

• Low: Fast responses for general dialogue.
• Medium: Balanced speed and detail.
• High: Deep and detailed analysis.

The reasoning level can be set in the system prompts, e.g., "Reasoning: high".

Tool use

The gpt-oss models are excellent for:

• Web browsing (using built-in browsing tools)
• Function calling with defined schemas
• Agentic operations like browser tasks

Fine-tuning

Both gpt-oss models can be fine-tuned for a variety of specialized use cases.

This smaller model gpt-oss-20b can be fine-tuned on consumer hardware, whereas the larger gpt-oss-120b can be fine-tuned on a single H100 node.

Citation

@misc{openai2025gptoss120bgptoss20bmodel,
      title={gpt-oss-120b & gpt-oss-20b Model Card}, 
      author={OpenAI},
      year={2025},
      eprint={2508.10925},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.10925}, 
}