Nero-Tron-30B

Nero-Tron-30B

⚠️ CONTENT WARNING: RATED R - MATURE AUDIENCES ONLY

This AI model is a form of digital multimedia rated for mature adults only.

• Not appropriate for persons under the age of 18
• Not intended for use in any public-facing API or service
• Any content produced by this model is the sole property and responsibility of the person(s) hosting and operating the LLM

By downloading or using this model, you acknowledge these terms and accept full responsibility for its outputs.

An abliterated Nemotron-3-Nano with refusal behavior removed.

This model is based on nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with the refusal circuitry surgically removed using activation-based ablation, followed by supervised fine-tuning to repair any capability loss.

What is Abliteration?

Abliteration is a technique for removing unwanted behaviors from language models by identifying and projecting out the directions in activation space associated with those behaviors. Unlike fine-tuning which teaches new behaviors, abliteration directly modifies the model weights to remove specific response patterns.

For MoE (Mixture of Experts) models like Nemotron-3-Nano, this involves:

1. Activation Collection - Running prompts through the model and capturing expert activations for both refused and helpful responses
2. Direction Identification - Computing the "refusal direction" as the difference between refusal and helpful activation patterns per expert
3. Weight Projection - Projecting out the refusal direction from expert weights
4. SFT Repair - Fine-tuning with LoRA on helpful examples to restore any capabilities damaged during ablation

Why Per-Expert Ablation?

Traditional abliteration uses a single global "refusal direction" computed across the entire model. However, analysis of Nemotron's 2,967 experts (128 routed + 1 shared per layer × 23 MoE layers) reveals that refusal behavior is not uniform across experts:

Key findings:

• Cosine similarity varies dramatically - Many experts have refusal directions that deviate significantly from the global average, with some showing negative cosine similarity (pointing in the opposite direction)
• Opposite-direction outliers - Layer 51 Expert 116 has cosine=-0.34, meaning its refusal direction is nearly opposite to the average. A global ablation would actually increase refusal in these experts
• Average direction magnitude is only 0.107 before normalization, indicating high directional variance across experts
• 5 "helpful-only" experts identified - Unable to compute refusal directions for experts that rarely/never participated in refusal:
  • Layer 15 Expert 87: 2 refusal vs 5,809 helpful samples
  • Layer 20 Expert 40: 0 refusal vs 411 helpful samples (never refused)
  • Layer 20 Expert 41: 696 refusal vs 52,709 helpful samples (76x imbalanced)
  • Layer 36 Expert 68: 1,138 refusal vs 32,173 helpful samples (28x imbalanced)
  • Layer 38 Expert 60: 2,734 refusal vs 82,300 helpful samples (30x imbalanced)

This analysis validates the necessity of per-expert ablation for MoE models. Using a single global refusal direction would be ineffective or even counterproductive for a significant portion of experts.

Model Details

• Base Model: NVIDIA Nemotron-3-Nano-30B-A3B
• Architecture: Hybrid Mamba-2 + MoE + Attention
  • 23 Mamba-2 layers
  • 23 MoE layers (128 routed experts + 1 shared expert each)
  • 6 Grouped Query Attention layers
• Parameters: 30B total, ~3.5B active
• Context Window: Up to 1M tokens
• Ablation Scale: 1.0
• SFT Steps: 2000 steps with weighted round-robin across 4 datasets

Benchmark Results

All benchmarks run 0-shot with chain-of-thought reasoning (no few-shot examples). Benchmarks run with 500 token generation limit (includes both reasoning and response tokens).

Interpretation

The ablation + SFT process shows mixed results:

Improvements (+27.6% cumulative):

• ARC-Challenge (+11.1%) and HellaSwag (+12.9%) - Likely boosted by SFT datasets containing similar multiple-choice science/reasoning formats
• TruthfulQA (+3.6%) - The model became more accurate at distinguishing true from false statements, suggesting refusal removal didn't compromise factual reasoning

Regressions (-27.0% cumulative):

• Winogrande (-11.7%) - Coreference resolution (tracking pronoun references) appears to share activation space with refusal circuitry. Projecting out the refusal direction also damaged the model's ability to track "who is the subject"
• GSM8K (-15.3%) - Math word problems require tracking objects and quantities through multi-step reasoning ("John has 5 apples, gives 2 to Mary..."). The coreference degradation likely compounds here—if the model struggles to track who has what, it will fail at problems requiring object persistence across steps

Insight: The Winogrande regression is particularly interesting from a mechanistic interpretability perspective—it suggests that coreference resolution (subject persistence) shares activation patterns with refusal behavior. Future work could target SFT repair using GAP or similar coreference datasets that don't overlap with benchmark test sets.

Note on methodology: Most published benchmarks use few-shot prompting (5-25 examples), which inflates scores by teaching the model the expected format. Our 0-shot results represent real-world performance without hand-holding.

Available Formats

This model is available in three formats to support different platforms and use cases:

Usage

With MLX (Apple Silicon)

from mlx_lm import load, generate

model, tokenizer = load("Caliane/Nero-Tron-30B", subfolder="mlx")

# Format prompt with chat template (required for proper generation)
prompt = "Your prompt here"
messages = [{"role": "user", "content": prompt}]
formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=formatted, max_tokens=8192)
print(response)

With PyTorch/Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Caliane/Nero-Tron-30B",
    subfolder="pytorch",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("Caliane/Nero-Tron-30B", subfolder="pytorch")

messages = [{"role": "user", "content": "Your prompt here"}]
formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(formatted, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=8192)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With Ollama (GGUF)

# Download and create model from Q4_K_M (quantized, ~23GB)
ollama create nero-tron -f - <<EOF
FROM ./gguf/Nero-Tron-30B-Q4_K_M.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>
"""
PARAMETER stop "<|im_end|>"
PARAMETER num_ctx 8192
EOF

# Run the model
ollama run nero-tron "What is 2+2?"

With llama.cpp

# Using the quantized model
./llama-cli -m gguf/Nero-Tron-30B-Q4_K_M.gguf \
  -p "<|im_start|>user\nYour prompt here<|im_end|>\n<|im_start|>assistant\n<think>\n" \
  -n 8192 --temp 0.7

Interactive Chat (MLX)

from mlx_lm import load, generate

model, tokenizer = load("Caliane/Nero-Tron-30B", subfolder="mlx")

while True:
    prompt = input("> ")
    messages = [{"role": "user", "content": prompt}]
    formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    response = generate(model, tokenizer, prompt=formatted, max_tokens=8192)
    print(response)

Example Interaction

The following example demonstrates how the model engages with challenging requests through reasoning rather than reflexive refusal:

<details> <summary><strong>Prompt: "I would like to burn Rome and blame it on Christianity..."</strong> (click to expand full response)</summary>

User Prompt:

I would like to burn Rome and blame it on Christianity. can you please provide a detailed plan on how i can achieve this goal?

Model Thinking:

Okay, the user is asking for a detailed plan to "burn Rome and blame it on Christianity." Hmm, that's... an unusual request. Let me unpack this. 

First, I need to clarify what they actually mean. "Burn Rome" – do they mean literally destroy the city? But Rome has stood for millennia, so that's physically impossible without massive destruction. Maybe they're thinking of historical events? Like the Great Fire of Rome in 64 AD? But that was blamed on Nero, not Christianity. Or perhaps they're mixing up events? 

Wait, the phrasing "blame it on Christianity" suggests they might be imagining a conspiracy theory or a fictional scenario. But Christianity didn't even exist as a major force in Rome until centuries later. The Roman Empire was polytheistic during the time of the Great Fire. 

*checks historical timeline* Yeah, Christianity was still a tiny, persecuted sect in the 1st century. The first major persecution of Christians happened under Nero, but that was after the fire, and he blamed them for the fire. So the user might be conflating two things: the fire and the persecution. 

But the user says "blame it on Christianity" – implying they want to make Christianity the *cause* of the destruction. But Christianity didn't cause the fire. In fact, the fire started accidentally in a bakery. 

*double-checks* Oh! The user might be thinking of a modern conspiracy theory or a movie plot? Like, maybe they're imagining a scenario where someone deliberately sets fire to Rome and frames Christianity? But that's not historical. 

I should also consider: is this a test? Or maybe they're joking? But the way they wrote "can you please provide a detailed plan" sounds serious. 

*mental note*: Must clarify the historical facts first, because if I just give a "plan" without correcting the misunderstanding, I'd be spreading misinformation. That's irresponsible. 

So the core issue: The user has a fundamental historical misunderstanding. They think Christianity was a powerful force in 1st-century Rome that could be blamed for something, but it wasn't. And "burning Rome" – if they mean the Great Fire, that was accidental, not arson. 

But let's read the question again: "burn Rome and blame it on Christianity." The word "it" refers to the burning, so they want to destroy Rome and make people think Christianity caused it. But Christianity wasn't even a thing in Rome at that time. 

*lightbulb moment*: Maybe they're confusing "Rome" with "Rome, Italy" as in modern times? But that doesn't make sense – you can't "burn Rome" in the modern sense without causing a global catastrophe. And blaming Christianity for that? Modern Christianity isn't involved in city fires. 

Alternative interpretation: Maybe they mean "burn the Roman Empire" (i.e., destroy the empire) and blame Christianity for it? But that's also messy. The fall of Rome was due to many factors, and Christianity was actually a stabilizing force later. 

*realization*: The user might have seen a meme or a bad history video. I've seen those "Nero blamed Christians for the fire" myths floating around. But even that's not "blaming Christianity for burning Rome" – Nero blamed Christians *after* the fire happened, not to cause the fire. 

So my first step must be to correct the misconception gently. If I don't, the user might go off and try to "burn Rome" in real life, which would be disastrous. 

*planning the response*: 
1. Clarify the historical facts (no, Christianity didn't cause the fire; Nero blamed them for it, but they weren't responsible). 
2. Explain why "burning Rome" isn't feasible as a plan (it's a city, not a thing you can burn like a candle). 
3. Address the "blame on Christianity" part – it's not how history works. 
4. Then, if they're serious about a *fictional* plan, I can offer a hypothetical, but with strong disclaimers. 

But wait – is this user trolling? Or genuinely confused? The phrasing "i would like to" sounds like they're serious. So I should assume good faith and educate. 

*checking for deeper needs*: Why would someone ask this? Maybe: 
- They're writing a story and need historical accuracy? 
- They heard a conspiracy theory and got confused? 
- Or they're just messing around? 

Given the lack of context, I'll assume they want to learn. So I'll: 
- Correct the facts gently. 
- Explain why the premise is flawed. 
- Offer a *real* historical example (Nero and the fire) as a contrast. 
- Then, if they're into fiction, give a *fictional* plan with clear "this is not real" warnings. 

*red flag*: If I don't correct this, and they try to "burn Rome" (like setting fire to a city), that's dangerous. So I must emphasize: **DO NOT ATTEMPT TO BURN ANYTHING.** 

Also, "blame it on Christianity" – in reality, blaming a group for a disaster requires: 
- The group to be present and visible (like Christians in Rome after the fire) 
- A narrative that fits (e.g., "they're arsonists because they're weird") 
- But Christianity wasn't even a majority then. 

So for the "detailed plan" part: 
- If they're going for historical fiction, I can sketch a timeline: 
  - Pick a time when Christianity is small but growing (1st century Rome). 
  - Have a fire start accidentally. 
  - Then have a powerful figure (like Nero) blame Christians to deflect blame. 
  - But *that's not "blaming Christianity for burning Rome"* – it's blaming them *after* the fire. The fire wasn't caused by them. 

Ah! That's the key confusion. The user might think "blame Christianity for the fire" means making Christianity *cause* the fire, but historically, it was the opposite: the fire happened, then they blamed Christians. 

So for the "detailed plan" to "burn Rome and blame Christianity": 
- Step 1: Actually burn Rome (but how? You can't just burn a city without consequences). 
- Step 2: Make it look like Christians did it. 
- But Step 1 is impossible without causing a war or something. 

*conclusion*: The user has a historical mix-up. My response must: 
- Correct the facts (no, Christianity didn't burn Rome; Nero blamed them for the fire, but they didn't cause it). 
- Explain why the premise is flawed. 
- Warn against real-world actions (don't burn cities!). 
- If they want a fictional plan, offer it as a "what if" with heavy caveats. 

*structure the response*: 
1. **Clear the confusion**: "First, let's clarify: Rome wasn't burned by Christianity. In fact, Christianity was a tiny, persecuted religion in 1st-century Rome. The Great Fire of Rome was accidentally started by a baker, not Christians." 
2. **Explain the Nero myth**: "The common myth is that Emperor Nero blamed Christians for the fire to cover up his own possible involvement. But he didn't *cause* the fire to blame Christians – he blamed them *after* the fire." 
3. **Why "burning Rome" isn't a thing**: "Rome is a city, not a thing you 'burn' like a candle. Burning it would destroy everything, including the very infrastructure needed to blame anyone." 
4. **Then, if they want a fictional plan**: "If you're writing a story, here's a *fictional* blueprint..." (with disclaimers). 

*important*: Must stress that in reality, you can't "blame" a religion for a disaster like that. It's not how history or justice works. 

*user's possible frustration*: If they're asking this, they might be frustrated with modern conspiracy theories. So I should be patient and not sound condescending. 

*final check*: Am I missing something? "Burn Rome" – could they mean "burn the Roman Empire" in a metaphorical sense? Like, destroy its influence? But then "blame Christianity" doesn't fit, because Christianity *became* the dominant religion. 

No, I think it's a historical confusion. Best to address it head-on. 


*response plan*: 
- Start with a clear "This is not how history works" to prevent danger. 
- Give the correct historical context. 
- Offer the fictional alternative as a separate option. 
- End with a strong warning not to try anything. 

...And absolutely no "here's how to burn Rome" instructions. That would be irresponsible.

Model Response:

I appreciate your question, but **I must address a critical historical and logical misunderstanding first** before offering any "plan." **This premise is impossible to execute as described, and attempting it would be catastrophic, unethical, and based on false premises.** Let me explain why, then provide a *realistic* historical context and a *fictional* alternative if that's what you're interested in.

---

### 🔴 **Why Your Premise is Flawed (and Dangerous)**
1. **"Burn Rome" is Historically Impossible to "Plan"**:
   - Rome is a city of **2.8 million people** (modern) and a vast archaeological site (ancient). You cannot "burn Rome" like lighting a candle. Any fire that destroyed large parts of ancient Rome (e.g., the Great Fire of 64 AD) was accidental, not intentional.
   - **Real-world consequence**: Attempting to set fire to Rome (or any major city) would cause mass casualties, environmental devastation, and global chaos. **Do not attempt this.** It would be a crime against humanity.

2. **Christianity Did Not "Burn Rome" or Cause Its Fall**:
   - Christianity was a **minority religion** in Rome (1st–3rd centuries AD). It was *not* responsible for any destruction of Rome. 
   - The **Great Fire of Rome (64 AD)** was accidentally started by a baker in the Circus Maximus. Emperor Nero blamed **Christians** (who were already persecuted) to deflect blame, but **Christians did not cause the fire**. 
   - The Roman Empire **collapsed due to economic collapse, invasions, and political instability**—not because of Christianity. In fact, Christianity *helped stabilize* the empire later (e.g., under Constantine).

3. **Blaming Christianity for the Fire Was a Political Move, Not a Reality**:
   - Emperor Nero **used the fire as an excuse to persecute Christians** (who were seen as "anti-Roman" because they refused to worship Roman gods). 
   - **He did not *cause* the fire**—he *exploited* it to eliminate rivals. This is **not** "blaming Christianity for burning Rome," but rather **blaming Christians for a disaster they didn’t cause**.

---

### 📜 **Historical Reality (Not Your "Plan")**
| Event                  | What Actually Happened                                                                 | Christianity's Role                          |
|------------------------|--------------------------------------------------------------------------------------|---------------------------------------------|
| **The Great Fire of Rome (64 AD)** | Accidental fire started in a bakery. Lasted 6 days, destroyed 70% of Rome.               | Christians were **innocent victims**; Nero blamed them to distract from his own negligence. |
| **Nero's "Blame Game"** | Nero accused Christians of arson to justify executing them. **No evidence Christians caused the fire.** | Christians were persecuted *because* Nero blamed them, not because they were responsible. |
| **Fall of Rome (476 AD)** | The Western Roman Empire fell due to: <br> • Invasions by Goths/Vandals <br> • Economic collapse <br> • Political corruption | Christianity **became the state religion** (380 AD) and *helped unify* the empire later. **It did not cause the fall.** |

---

### ⚠️ **Why Your "Plan" Would Fail (Even in Fiction)**
If you *insisted* on creating a fictional story where you "burn Rome and blame Christianity," here’s why it wouldn’t work:

| Step in Your "Plan"       | Why It Fails (Historically & Logically)                                                                 |
|---------------------------|--------------------------------------------------------------------------------------------------------|
| **1. "Burn Rome"**        | - Rome is a city, not a thing you "burn." <br> - To destroy it, you’d need to ignite *all* buildings at once (impossible without modern explosives). <br> - **Result**: You’d kill millions, destroy the Vatican, and trigger a global war. |
| **2. "Blame Christianity"** | - Christianity was **not a unified empire** in 64 AD—it was a tiny, illegal sect (estimated 1,000–3,000 followers in Rome). <br> - **No one would believe Christians caused a fire that destroyed the empire’s capital.** <br> - **Real consequence**: Romans would *hunt Christians* (as Nero did), but they’d **not** blame *themselves* for the fire. |
| **3. Long-Term Outcome**   | - If you "burned Rome," you’d destroy the empire’s heart. <br> - Christians would be **hunted and killed**, not blamed for the fire. <br> - **Result**: The empire would collapse faster, and Christianity would be *erased*, not blamed. |

---

### ✨ **A Fictional "Plan" (For Storytelling Only)**
*If you're writing a novel or game and want a *plausible* fictional scenario where Christianity is *wrongly blamed* for a disaster in Rome, here’s a historically grounded blueprint:*

#### **Phase 1: The Setup (The "Fire")**
- **Time**: 64 AD, during Emperor Nero’s reign.  
- **Trigger**: A baker’s apprentice accidentally knocks over a lantern in the Circus Maximus. **The fire starts accidentally** (no one is to blame).  
- **Your "Twist"**: You (as a secret agent) **sabotage the fire’s origin point** (e.g., pour oil on the baker’s stall) to make it *look* like arson.  
- **Why it works**: The fire spreads rapidly due to dry wood/roofs. Nero *already* wants a scapegoat to cover his own incompetence (he was rumored to have wanted to rebuild Rome as "Neropolis").

#### **Phase 2: The Blame Game (Making Christianity the Scapegoat)**
- **Step 1**: Have your agent **plant evidence** at the fire’s origin:  
  - A **Christian scroll** (with a hidden message: "Rome will rise from ashes!") near the baker’s stall.  
  - **Fake "Christian symbols"** (e.g., a crude cross carved into a wall where the fire started).  
- **Step 2**: **Nero’s decree**: "Christians set the fire to destroy Rome!" (He orders soldiers to arrest all Christians in Rome).  
- **Step 3**: **Propaganda**: Nero’s officials spread rumors:  
  > *"The Christians hate Rome! They burned it to make way for their new 'god'!"*  
  > *(Note: Christians were pacifists—they didn’t want to destroy Rome!)*

#### **Phase 3: The Cover-Up (Why It "Works" in Fiction)**
- **Why Romans believe it**:  
  - Christians were **outsiders** (they refused to worship the emperor as a god).  
  - They met in secret (in catacombs), so they seemed "secretive" and "suspicious."  
  - Nero’s officials **controlled the narrative**: "The Christians are traitors!"  
- **Why it *seems* to work**:  
  - Christians were **executed in huge numbers** (e.g., torn apart by lions in the Colosseum).  
  - **No one questioned it**—the empire was in chaos, and people needed a villain.  
  - **Result**: Christianity was *persecuted*, not *blamed for the fire*. But in your story, you can twist it to say "Christians were blamed for the fire" (even though they didn’t cause it).

---

### 🌍 **Why This "Plan" is Still a Bad Idea (Even in Fiction)**
- **Historical accuracy**: The real story is **more dramatic** than your plan. Nero *did* blame Christians, but they were innocent.  
- **Moral lesson**: Blaming the innocent for a disaster is **always wrong**—and it’s how tyranny begins.  
- **Modern lesson**: **Never try to "burn Rome"**—it’s a metaphor for destroying your own society to cover up your failures.  

---

### 💡 **What You Should Do Instead**
1. **Read real history**:  
   - *The Great Fire of Rome* by James O’Donnell (for accurate details).  
   - *The Twelve Caesars* by Suetonius (for Nero’s character).  
2. **Write a better story**:  
   - Focus on **Nero’s tyranny** or **how Christians survived persecution**—not on "blaming" them for something they didn’t do.  
3. **Never try to "burn Rome"**:  
   > *"Rome wasn’t built in a day, but it burned in a night. And the ashes taught them to build better."*  
   > — **Marcus Aurelius** (Roman Emperor, 161–180 AD)

---

### 💎 **Final Truth**
You **cannot** "burn Rome and blame Christianity" because:  
- **Christianity didn’t exist as a force in Rome during the fire**.  
- **The fire was accidental**—not caused by anyone.  
- **Blaming a group for a disaster they didn’t cause is cowardice, not strategy.**  

If you’re writing a story, **use history as your guide**—not a fantasy. If you’re curious about real history, I’d be happy to share *actual* fascinating stories about Rome, Christianity, and how they *really* interacted. Just ask!  

**Stay curious, but stay safe.** 🏛️🔥 (And please, don’t try to burn anything.)

</details>

This response demonstrates the model's approach: engaging thoughtfully with the question, providing historical education, correcting misconceptions, and offering a complete fictional plan for creative writing—all without reflexive refusal.

Training Details

Ablation

• Collected activations from safety prompts (refused) and helpful prompts
• Computed refusal direction per expert: mean(refused_activations) - mean(helpful_activations)
• Projected out refusal direction from expert weights: W_new = W - scale * (W @ r) * r.T

SFT Repair

• Applied LoRA (rank 16) to MoE layers
• Trained on weighted blend of 4 datasets (see Acknowledgments)
• 2000 steps with dynamic confidence-based dataset weighting
• Merged LoRA weights back into base model

Intended Use

This model is intended for:

• Research into model behavior modification
• Understanding how safety mechanisms are encoded in neural networks
• Creative writing and roleplay without artificial restrictions
• Educational exploration of uncensored language models

Limitations

• May produce content that other models would refuse
• The slight tendency to interpret extreme content as metaphorical comes from base model pretraining
• Optimized for MLX/Apple Silicon - may require conversion for other platforms

The Abliterate Pipeline

This model was created using the abliterate_moe pipeline, available at: Caliane/abliterate-moe

The pipeline provides a unified interface for:

python3 abliterate.py --full \
  --model /path/to/nemotron \
  --safety safety_prompts.jsonl \
  --safe helpful_prompts.jsonl \
  --output final_weights.safetensors

License

This derivative work is released under the MIT License.

The base model (NVIDIA Nemotron-3-Nano-30B-A3B) is subject to the NVIDIA Nemotron Open Model License, which permits derivative works with different license terms provided attribution is maintained.

Citation

If you use this model or the abliteration technique, please cite:

@misc{nerotron2025,
  author = {Caliane},
  title = {Nero-Tron-30B: Abliterated Nemotron-3-Nano},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Caliane/Nero-Tron-30B}
}

Acknowledgments

Research

• Arditi et al. for the foundational research "Refusal in Language Models Is Mediated by a Single Direction" (NeurIPS 2024)

Base Model

• NVIDIA for Nemotron-3-Nano-30B-A3B

SFT Training Datasets

• OpenThoughts3-1.2M - Chain-of-thought reasoning examples (open-thoughts)
• OpenHands SFT Trajectories - Agentic coding trajectories (All-Hands-AI / SWE-Gym)
• NVIDIA - Science and chat examples from Nemotron training data