text string | quality_score float64 | source string | row_hash string | first_source_dataset string | first_source_config string | first_source_split string | first_source_row_index int64 | seen_count int64 | row_json string | id string | category string | difficulty string | instruction string | input string | output string | tags list | response string | metadata dict | thinking_trace string | query string | thinking string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
id: seedai_100798 | category: self_assessment | difficulty: expert | instruction: You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-Coder-32B" on the metric "DocVQA" (current estimated score: 91%).
Your task:
1. Break down the model's likely strengths and weak... | null | null | null | null | null | null | null | null | null | seedai_100798 | self_assessment | expert | You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-Coder-32B" on the metric "DocVQA" (current estimated score: 91%).
Your task:
1. Break down the model's likely strengths and weaknesses in "code generation & debugging" using real evaluation frameworks.
2. Ident... | Current model: Qwen2.5-Coder-32B
Target metric: DocVQA
Current estimated performance: 91%
Primary domain focus: code generation & debugging
Available compute: 8xH100 (self-hosted) | <think>
**Step 1: Rigorous Self-Assessment (Fact-Based)**
Current estimated performance on DocVQA: 91%
**Strengths (based on architecture and training data patterns):**
- Strong performance on structured reasoning tasks due to post-training alignment (similar to Llama-3.3 and Qwen2.5 results).
- Good tool-calling whe... | [
"recursive",
"self-assessment",
"seed_ai",
"code_generation_&_debugging",
"think_tags"
] | null | null | null | null | null |
<|user|>
Develop a backtracking algorithm to solve a Sudoku puzzle efficiently.
<|assistant|>
<think>First, the user asked to develop a backtracking algorithm to solve a Sudoku puzzle efficiently. This is an open-ended question, so I need to make my response maximally informative. Structure it easily parsable, perhaps ... | 0.85 | grok_4.4_distilled | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
<|user|>
How might we validate ambiguous or open-ended requirements in safety-critical systems like medical devices or autonomous vehicles?
<|assistant|>
<think>**Exploring validation methods**
The user is asking about validating ambiguous or open-ended requirements in safety-critical systems, like medical devices o... | 0.85 | gpt_5.5_distilled | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
id: seedai_637388 | category: training_recipe | difficulty: expert | instruction: Design a complete, fact-based training recipe to turn Llama-3.1-405B into a stronger Recursive Seed AI focused on "achieve superhuman performance on GPQA" with measurable gains on HumanEval+.
Requirements:
- Use only published, reproduc... | null | null | null | null | null | null | null | null | null | seedai_637388 | training_recipe | expert | Design a complete, fact-based training recipe to turn Llama-3.1-405B into a stronger Recursive Seed AI focused on "achieve superhuman performance on GPQA" with measurable gains on HumanEval+.
Requirements:
- Use only published, reproducible techniques.
- Specify exact hyperparameters, data mixture ratios, compute bud... | Base model: Llama-3.1-405B
Primary goal: achieve superhuman performance on GPQA
Target benchmark: HumanEval+
Hardware: 8xH100 80GB
Current estimated performance: 71% | <think>
**Complete Training Recipe: Recursive Seed AI v2 (from Llama-3.1-405B)**
**Phase 0: Baseline & Infrastructure (1 day)**
- Quantize base model to 4-bit NF4 using bitsandbytes (QLoRA setup from Dettmers et al. 2023).
- Set up evaluation harness with HumanEval+ + 3 auxiliary metrics.
- Create 2,000 high-quality s... | [
"training",
"recipe",
"LoRA",
"recursive_data",
"HumanEval+",
"think_tags"
] | null | null | null | null | null |
system: You are an expert software engineer. Write complete, correct, runnable code.
Format your response EXACTLY as:
<think>
[Analyze the problem, consider edge cases, choose your approach,
explain your algorithm and data structure choices.]
</think>
```python
[Complete, runnable solution with docstring and type hin... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
row_hash: 5f960b7fd067afd94eb1d8e2b89b299d23a0ffde2f0a5787a9d14dc4cf8032be | first_source_dataset: 1EYE4ALL/Fable-5-traces | first_source_config: local_files | first_source_split: claude/projects/-home-/68396a0d-8585-4d00-9321-54b2b00c07f4.jsonl | first_source_row_index: 523 | seen_count: 4 | row_json: {"cwd":"/home/la... | null | null | 5f960b7fd067afd94eb1d8e2b89b299d23a0ffde2f0a5787a9d14dc4cf8032be | 1EYE4ALL/Fable-5-traces | local_files | claude/projects/-home-/68396a0d-8585-4d00-9321-54b2b00c07f4.jsonl | 523 | 4 | {"cwd":"/home/lane/.paperclip/adapter-plugins","entrypoint":"cli","gitBranch":"HEAD","isSidechain":false,"message":{"content":[{"signature":"CAISgQQKYggOGAIqQHz8wzzL8bPCUF0b6QSVidSY9e+kv+kIveAfnkIfCjsIXyKA8iMcvMADit+0i6pdz8FWUO+30IAYLnJhrMd9vT0yDmNsYXVkZS1mYWJsZS01OAFCCHRoaW5raW5nEgz+g0MEez1PnFsxM2oaDBUXMnARGvM7z8IFNCI... | null | null | null | null | null | null | null | null | null | null | null | null |
query: Design a distributed bloom filter system for 1PB with causal replication. | thinking: <think>
Alright, I need to think through this carefully.
The user wants to: Design a distributed bloom filter system for 1PB with causal replication. Let me analyze what this entails.
Let me think about the disaster recovery ... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | Design a distributed bloom filter system for 1PB with causal replication. | <think>
Alright, I need to think through this carefully.
The user wants to: Design a distributed bloom filter system for 1PB with causal replication. Let me analyze what this entails.
Let me think about the disaster recovery strategy. RTO (Recovery Time Objective) determines how quickly we need to restore service. RP... |
user: === MOONSHINER TASK BOUNDARY ===
Our depot's cold-chain recorder library stopped building the moment we switched it onto the hardened CI profile (optimized + fortified, warnings fatal — the Makefile in the repo is exactly what CI runs). The reader and summary logic have been in the vans all season and the number... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
id: seedai_405279 | category: self_assessment | difficulty: expert | instruction: You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "HumanEval+" (current estimated score: 78%).
Your task:
1. Break down the model's likely strengths a... | null | null | null | null | null | null | null | null | null | seedai_405279 | self_assessment | expert | You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "HumanEval+" (current estimated score: 78%).
Your task:
1. Break down the model's likely strengths and weaknesses in "continual learning" using real evaluation frameworks.
2. Identif... | Current model: Qwen2.5-72B-Instruct
Target metric: HumanEval+
Current estimated performance: 78%
Primary domain focus: continual learning
Available compute: 8xH100 (self-hosted) | <think>
**Step 1: Rigorous Self-Assessment (Fact-Based)**
Current estimated performance on HumanEval+: 78%
**Strengths (based on architecture and training data patterns):**
- Strong performance on structured reasoning tasks due to post-training alignment (similar to Llama-3.3 and Qwen2.5 results).
- Good tool-calling... | [
"recursive",
"self-assessment",
"seed_ai",
"continual_learning",
"think_tags"
] | null | null | null | null | null |
instruction: Generate a valid THRIFT API payload for `deleteAlert` using Redis Pub/Sub with the following constraints:
- Endpoint parameters: {"name": "conditional", "sort_by": "optional", "region": "optional", "timeout": "required", "tags": "required"}
- Max retries: 4
- Schema version: 4.6
- Response format: THRIFT |... | null | null | null | null | null | null | null | null | null | null | null | null | Generate a valid THRIFT API payload for `deleteAlert` using Redis Pub/Sub with the following constraints:
- Endpoint parameters: {"name": "conditional", "sort_by": "optional", "region": "optional", "timeout": "required", "tags": "required"}
- Max retries: 4
- Schema version: 4.6
- Response format: THRIFT | null | <thought>
1. Verify payload passes strict schema validation before sending.
2. Validate that all required fields from {"name": "conditional", "sort_by": "optional", "region": "optional", "timeout": "required", "tags": "required"} are present.
3. Identify required vs optional parameters for deleteAlert.
</thought>
```
... | null | null | {
"category": "systemic_execution_instructions",
"index": 340
} | null | null | null |
system: You are a helpful, precise assistant. Follow instructions exactly.
Format your response as:
<think>
[Understand what is being asked. Identify constraints and requirements.
Plan your response structure.]
</think>
[Your complete, well-structured response following all given constraints]
user: Yes / no, is the f... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
id: seedai_587290 | category: self_assessment | difficulty: expert | instruction: You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "GSM8K" (current estimated score: 86%).
Your task:
1. Break down the model's likely strengths and we... | null | null | null | null | null | null | null | null | null | seedai_587290 | self_assessment | expert | You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "GSM8K" (current estimated score: 86%).
Your task:
1. Break down the model's likely strengths and weaknesses in "long-context understanding" using real evaluation frameworks.
2. Iden... | Current model: Qwen2.5-72B-Instruct
Target metric: GSM8K
Current estimated performance: 86%
Primary domain focus: long-context understanding
Available compute: 8xH100 (self-hosted) | **Step 1: Rigorous Self-Assessment (Fact-Based)**
Current estimated performance on GSM8K: 86%
**Strengths (based on architecture and training data patterns):**
- Strong performance on structured reasoning tasks due to post-training alignment (similar to Llama-3.3 and Qwen2.5 results).
- Good tool-calling when prompts... | [
"recursive",
"self-assessment",
"seed_ai",
"long-context_understanding"
] | null | null | null | null | null |
row_hash: 81c6c4bc8760e86d10155b04fd720ea952198a53b395222a00d6de3446447207 | first_source_dataset: lordx64/agentic-distill-fable-5-sft | first_source_config: default | first_source_split: train | first_source_row_index: 3855 | seen_count: 1 | row_json: {"session":"glint","source":"Glint-Research/Fable-5-traces","text":... | null | null | 81c6c4bc8760e86d10155b04fd720ea952198a53b395222a00d6de3446447207 | lordx64/agentic-distill-fable-5-sft | default | train | 3,855 | 1 | {"session":"glint","source":"Glint-Research/Fable-5-traces","text":"<|im_start|>system\nYou are a helpful AI assistant.<|im_end|>\n<|im_start|>user\n…[earlier truncated]…\n13\t)\n14\tfrom app.agents.title_extraction import empty_extraction, normalize_extraction\n15\t\n16\t\n17\tdef _wrap(v, c=0.9):\n18\t return {\"v... | null | null | null | null | null | null | null | null | null | null | null | null |
row_hash: 4f66ae91fb60977868009bf77ad2dde7c057ef3a82ffb16104127ec2c3902550 | first_source_dataset: lordx64/agentic-distill-fable-5-sft | first_source_config: default | first_source_split: train | first_source_row_index: 397 | seen_count: 1 | row_json: {"session":"glint","source":"Glint-Research/Fable-5-traces","text":"... | null | null | 4f66ae91fb60977868009bf77ad2dde7c057ef3a82ffb16104127ec2c3902550 | lordx64/agentic-distill-fable-5-sft | default | train | 397 | 1 | {"session":"glint","source":"Glint-Research/Fable-5-traces","text":"<|im_start|>system\nYou are a helpful AI assistant.<|im_end|>\n<|im_start|>user\n…[earlier truncated]…\nnt in your context — no need to Read it back)\nASSISTANT (tool call) Edit input={\"replace_all\": false, \"file_path\": \"/home/lane/Blindbot/hf_spa... | null | null | null | null | null | null | null | null | null | null | null | null |
<|user|>
Design a spacecraft trajectory to Mars using gravitational assists.
<|assistant|>
Let me start by asking the right questions.
I'm also wondering about:
Let me explore different angles:
Let me think about what experimental evidence or observations we have.
Let me test my assumptions:
Based on what I've lea... | 0.5 | gpt_5.5_distilled | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
id: seedai_390883 | category: training_recipe | difficulty: expert | instruction: Design a complete, fact-based training recipe to turn Mistral-Small-3.1 into a stronger Recursive Seed AI focused on "create self-sustaining improvement loop" with measurable gains on MMLU.
Requirements:
- Use only published, reproducib... | null | null | null | null | null | null | null | null | null | seedai_390883 | training_recipe | expert | Design a complete, fact-based training recipe to turn Mistral-Small-3.1 into a stronger Recursive Seed AI focused on "create self-sustaining improvement loop" with measurable gains on MMLU.
Requirements:
- Use only published, reproducible techniques.
- Specify exact hyperparameters, data mixture ratios, compute budge... | Base model: Mistral-Small-3.1
Primary goal: create self-sustaining improvement loop
Target benchmark: MMLU
Hardware: 8xH100 80GB
Current estimated performance: 71% | **Complete Training Recipe: Recursive Seed AI v2 (from Mistral-Small-3.1)**
**Phase 0: Baseline & Infrastructure (1 day)**
- Quantize base model to 4-bit NF4 using bitsandbytes (QLoRA setup from Dettmers et al. 2023).
- Set up evaluation harness with MMLU + 3 auxiliary metrics.
- Create 2,000 high-quality seed example... | [
"training",
"recipe",
"LoRA",
"recursive_data",
"MMLU"
] | null | null | null | null | null |
query: Design low-power protocol for 11 devices with long range. | thinking: <think>
This requires careful analysis. Let me break it down.
The user wants to: Design low-power protocol for 11 devices with long range. Let me analyze what this entails.
Low-power communication protocols balance range, data rate, and powe... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | Design low-power protocol for 11 devices with long range. | <think>
This requires careful analysis. Let me break it down.
The user wants to: Design low-power protocol for 11 devices with long range. Let me analyze what this entails.
Low-power communication protocols balance range, data rate, and power consumption. LoRaWAN provides long range (kilometers) at low data rates (kb... |
id: seedai_706395 | category: self_assessment | difficulty: expert | instruction: You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "OSWorld" (current estimated score: 70%).
Your task:
1. Break down the model's likely strengths and ... | null | null | null | null | null | null | null | null | null | seedai_706395 | self_assessment | expert | You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "OSWorld" (current estimated score: 70%).
Your task:
1. Break down the model's likely strengths and weaknesses in "prompt engineering" using real evaluation frameworks.
2. Identify t... | Current model: Qwen2.5-72B-Instruct
Target metric: OSWorld
Current estimated performance: 70%
Primary domain focus: prompt engineering
Available compute: 8xH100 (self-hosted) | **Step 1: Rigorous Self-Assessment (Fact-Based)**
Current estimated performance on OSWorld: 70%
**Strengths (based on architecture and training data patterns):**
- Strong performance on structured reasoning tasks due to post-training alignment (similar to Llama-3.3 and Qwen2.5 results).
- Good tool-calling when promp... | [
"recursive",
"self-assessment",
"seed_ai",
"prompt_engineering"
] | null | null | null | null | null |
system: You are an expert logical reasoner. Think through the problem carefully.
Format your response EXACTLY as:
<think>
[Break down the problem into components.
Consider multiple hypotheses if applicable.
Eliminate incorrect paths explicitly.
Build toward the conclusion step by step.]
</think>
[Clear, concise final... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
<|user|>
Our team has been maintaining a gravitational coupling index between microservices for the past year — it measures how strongly each service 'attracts' traffic from adjacent services. The index keeps drifting. Should we recalibrate quarterly or switch to a continuous measurement?
<|assistant|>
# Honest respons... | 0.8 | opus_4.7_distilled | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
<|user|>
Casey has 68 pencils. Casey then discovers 14 more pencils. How many pencils does Casey have now?
<|assistant|>
To give the most reliable answer, I'll first restate the problem in my own words and then solve it step by step.
Beginning with 68, the operation described adds/subtracts 14. Because the items are f... | 0.65 | opus_4.7_distilled | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
<|user|>
Explain Tomasulo's algorithm for out-of-order execution. Describe reservation stations and common data bus. How does register renaming solve WAW and WAR hazards?
<|assistant|>
<think>**Explaining Tomasulo's algorithm**
Tomasulo's algorithm enables out-of-order execution via dynamic scheduling using reservatio... | 0.9 | gpt_5.5_distilled | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
<|user|>
Oscar has 31 buttons. Oscar then receives as a gift 8 more buttons. How many buttons does Oscar have now?
<|assistant|>
Let's approach this methodically to avoid any oversight. After identifying the starting amount and the incremental change, the mathematics is elementary but I still verify the arithmetic twic... | 0.65 | opus_4.7_distilled | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
system: You are an expert mathematician. Solve the given problem with complete rigor.
Format your response EXACTLY as:
<think>
[Your complete step-by-step reasoning. Show all intermediate calculations.
If you make an error and catch it, show the correction.
End with a clear statement of the final answer.]
</think>
\b... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
row_hash: c8f38309214be72b101e722a8f385795acada1c53119b337713b5e8d9da07215 | first_source_dataset: lordx64/agentic-distill-fable-5-sft | first_source_config: default | first_source_split: train | first_source_row_index: 112 | seen_count: 1 | row_json: {"session":"glint","source":"Glint-Research/Fable-5-traces","text":"... | null | null | c8f38309214be72b101e722a8f385795acada1c53119b337713b5e8d9da07215 | lordx64/agentic-distill-fable-5-sft | default | train | 112 | 1 | {"session":"glint","source":"Glint-Research/Fable-5-traces","text":"<|im_start|>system\nYou are a helpful AI assistant.<|im_end|>\n<|im_start|>user\n…[earlier truncated]…\n\"Run multiplayer debug from test dir with deps\", \"timeout\": 120000}\nTOOL RESULT: A internals {\n mode: 'dm',\n st: 'buy',\n room: 'NZ2H',\n ... | null | null | null | null | null | null | null | null | null | null | null | null |
query: Implement real-time control for event sourcing with <100ms latency. | thinking: <think>
I need to reason through this methodically.
The user wants to: Implement real-time control for event sourcing with <100ms latency. Let me analyze what this entails.
OTA firmware updates require careful design to avoid brick... | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | Implement real-time control for event sourcing with <100ms latency. | <think>
I need to reason through this methodically.
The user wants to: Implement real-time control for event sourcing with <100ms latency. Let me analyze what this entails.
OTA firmware updates require careful design to avoid bricking devices. The update image must be signed to prevent malicious firmware. A/B partiti... |
id: seedai_389569 | category: self_assessment | difficulty: expert | instruction: You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "GSM8K" (current estimated score: 71%).
Your task:
1. Break down the model's likely strengths and we... | null | null | null | null | null | null | null | null | null | seedai_389569 | self_assessment | expert | You are acting as a Recursive Seed AI. Perform a rigorous, fact-based self-assessment of the model "Qwen2.5-72B-Instruct" on the metric "GSM8K" (current estimated score: 71%).
Your task:
1. Break down the model's likely strengths and weaknesses in "advanced reasoning" using real evaluation frameworks.
2. Identify the... | Current model: Qwen2.5-72B-Instruct
Target metric: GSM8K
Current estimated performance: 71%
Primary domain focus: advanced reasoning
Available compute: 8xH100 (self-hosted) | <think>
**Step 1: Rigorous Self-Assessment (Fact-Based)**
Current estimated performance on GSM8K: 71%
**Strengths (based on architecture and training data patterns):**
- Strong performance on structured reasoning tasks due to post-training alignment (similar to Llama-3.3 and Qwen2.5 results).
- Good tool-calling when... | [
"recursive",
"self-assessment",
"seed_ai",
"advanced_reasoning",
"think_tags"
] | null | null | null | null | null |
End of preview. Expand in Data Studio
Frontier_Traces_Pre_Train
Massive merged & deduplicated frontier model distillation traces for pre-training from scratch.
Dataset Details
- Size: 316,299 unique examples
- Format: Single
textcolumn (all conversations normalized) - Deduplicated: Yes (MD5 hash on full text)
- Sources:
- GPT-5.x series
- Claude Opus/Mythos
- Grok 4.x
- Gemini 3.x / Pro
- Qwen
- DeepMind Alpha
- Fable traces
- And many more high-quality distilled reasoning & coding traces
Usage
from datasets import load_dataset
dataset = load_dataset("yourusername/Frontier_Traces_Pre_Train", split="train")
print(dataset[0])
- Downloads last month
- 96