instruction stringlengths 6 89 | input stringlengths 174 316 | output stringlengths 50 512 |
|---|---|---|
What does curatorKIT do? | CuratorKIT builds trainer-ready datasets for LLM post-training. Every generated answer is checked against the exact source chunk it came from. Rejected samples are written to rejected.jsonl with a structured reason, never dropped silently. A run also writes manifest.json, dataset_card.md, and SHA-256 checksums.txt. | CuratorKIT builds trainer-ready datasets for LLM post-training. |
What are the scoring mechanisms for Quality Gates? | Quality gates run after generation. HallucinationGate scores grounding against the source passage. RewardGate scores helpfulness, honesty, and instruction following. DiversityGate drops near-duplicates using sentence embeddings. Adaptive recovery diagnoses each rejection and repairs the recoverable ones. | Quality gates score grounding against the source passage with HallucinationGate, which evaluates whether the generated content is grounded in reality; RewardGate assesses the effectiveness of the generated content through helpfulness, honesty, and instructions, providing guidance on how to follow these guidelines; Dive... |
How does Adaptive Recovery work? | Quality gates run after generation. HallucinationGate scores grounding against the source passage. RewardGate scores helpfulness, honesty, and instruction following. DiversityGate drops near-duplicates using sentence embeddings. Adaptive recovery diagnoses each rejection and repairs the recoverable ones. | Adaptive Recovery diagnoses each rejection and repairs the recoverable ones, ensuring that the generated content meets all required criteria. |
": "Wh | Hygiene stages run before any LLM call. SecretsGate rejects API keys and private tokens. PIIPseudonymizer replaces names and emails with consistent fake values. ToxicityGate uses a local Detoxify classifier, with an optional LLM judge for borderline rows. | t are the functions of each component mentioned in the given text?",
"answer": "The components mentioned in the text include:
- SecretGate: Rejects API keys and private tokens to ensure data privacy.
- PIIPseudonymizer: Replaces names and emails with consistent fake values to enhance security.
- ToxicityGate: ... |
What are the steps involved in running a curator? | A Curator run is config then execute. CuratorConfig names the dataset, field mapping, generation task, and export formats. curator.run() walks readers, hygiene, generation, gates, and exporters in that order. Changing BACKEND does not require rewriting the curator cell. | Running a curator involves several key components: naming the dataset, field mapping, generating tasks, and exporting formats. The curator runs these operations in the specified order. |
How does changing the backend affect curator functionality? | A Curator run is config then execute. CuratorConfig names the dataset, field mapping, generation task, and export formats. curator.run() walks readers, hygiene, generation, gates, and exporters in that order. Changing BACKEND does not require rewriting the curator cell. | Changing the backend does not directly affect the functionality of the curator. However, if the backend changes, it may need to rewrite the curator code or adjust configurations accordingly. |
What technology does resolve_backend use for backend deployment? | resolve_backend picks OpenRouter, a local vLLM OpenAI server, or Ollama. vLLM install uses uv pip install vllm --torch-backend=cu128. The waiter polls /v1/models and does not kill a serve that is still loading weights. | resolve_backend selects OpenRouter as a local vLLM (Viable Large Language Model) OpenAI server or Ollama. |
Does Ollama use cu128 as the CUDA implementation backend? | resolve_backend picks OpenRouter, a local vLLM OpenAI server, or Ollama. vLLM install uses uv pip install vllm --torch-backend=cu128. The waiter polls /v1/models and does not kill a serve that is still loading weights. | Ollama employs cu128 as the CUDA implementation backend to ensure efficient execution of the model. |
What are the different types of JSONL readers? | JSONL readers accept Alpaca, ShareGPT, DPO, GRPO, and pretrain layouts. field_mapping remaps source columns onto schema fields. format='pretrain' keeps a text column as language-modeling output instead of renaming it to instruction. | JSONL readers can accept various models such as Alpaca, ShareGPT, DPO, GRPO, and pretrain layouts. They also support field mapping that renames source columns to corresponding schema fields. |
What is the main topic of the passage? | QA generation asks a grounded question per chunk and writes an answer that must be supported by that chunk. num_questions controls how many pairs each passage yields. difficulty is easy, medium, or hard and only changes the prompt, not the schema. | The passage discusses the requirements for generating QA questions. |
How does the author define 'grounded' questions? | QA generation asks a grounded question per chunk and writes an answer that must be supported by that chunk. num_questions controls how many pairs each passage yields. difficulty is easy, medium, or hard and only changes the prompt, not the schema. | The author defines grounded questions as those that are factually based and have clear connections to the content being asked about. |
What is the purpose of preference generation? | Preference generation writes a chosen completion and a rejected completion for the same prompt. single_call asks the model for both in one response. two_pass samples independently then labels the better one. Export is DPO JSONL. | Preference generation is used to assign a preference score or ranking to each option before deciding which one to use as the final choice. |
What does GRPO generation do? | GRPO generation samples several completions per prompt and optional judge scores. Temperature spread varies rollouts around llm_temperature so the group is not identical. Export includes grpo and ppo JSONL. | GRPO generation samples several completions per prompt and optional judge scores. It also spreads temperature across rollout ranges to ensure that the group is not identical. |
How does the export include grpo and ppo JSONL? | GRPO generation samples several completions per prompt and optional judge scores. Temperature spread varies rollouts around llm_temperature so the group is not identical. Export includes grpo and ppo JSONL. | The export includes both grpo and ppo JSONL files. |
What is the purpose of multi-turn generation? | Multi-turn generation builds a short dialogue from a source chunk. num_turns is the number of user-assistant pairs. ShareGPT is the natural export; Alpaca flattens the last turn. | Multi-turn generation is a method used to create a concise and coherent conversation between two users or assistants based on a given source chunk. It involves breaking down the input into multiple turns where each turn can represent a separate statement or query. |
How does shareGAN work? | Multi-turn generation builds a short dialogue from a source chunk. num_turns is the number of user-assistant pairs. ShareGPT is the natural export; Alpaca flattens the last turn. | ShareGAN is a technique that uses a neural network to output a sequence of words or phrases that describe something in a way that closely resembles the original text. This allows for more flexible and contextually appropriate responses compared to traditional language models. |
What does Evol-Instruct do? | Evol-Instruct rewrites a seed instruction into a harder variant. num_evolutions is how many rewrite steps run. generate_answers asks the model to solve the evolved prompt so the row is SFT-ready. | Evol-Instruct rewrites a seed instruction into a harder variant and generates answers for the generated prompts. |
": "Wh | Chain-of-thought generation either wraps an existing answer with reasoning or produces a new reasoned answer from the chunk. cot_mode is generate or wrap. The reasoning string is stored on the sample before export. | t does 'cot_mode' represent in the context of Chain-of-thought generation?",
"answer": "Cot_mode" refers to the mode of operation for chain-of-thought generation."
},
{
"question": "How does the reasoning string for 'cot_mode' differ when it's wrapped versus generated?",
"answer": "When wrapping, the re... |
What does the 'adversarial QA' conceptually entail? | Adversarial QA injects a failure into a subset of answers after generation. injection_rate is the fraction of rows that get a known bad span. injection_types include contradicts_source and parametric_drift. The metadata flag injected_failure marks those rows. | The 'adversarial QA' conceptually involves injecting a failure into a subset of answers generated during the process. This failure can occur when certain types of questions or parameters are used incorrectly, leading to unintended results. |
What is the purpose of HallucinationGate? | HallucinationGate scores whether the answer is entailed by the source passage. hallucination_threshold drops rows below the cutoff. A separate judge model avoids self-scoring. skip_if_no_context leaves rows without a passage alone. | HallucinationGate evaluates whether the answer to a question is supported by the information contained within the source passage. |
What is the purpose of the 'reward_threshold' parameter in the RewardGate scoring system? | RewardGate scores helpfulness, honesty, and instruction following. reward_threshold drops low-scoring rows. Scores can be stored on the label for later filtering. The judge uses the same API base as generation unless overridden. | The 'reward_threshold' parameter in the RewardGate scoring system drops low-scoring rows, allowing for the storage of scores on the label for further filtering. |
What does 'embedding model' refer to in this context? | DiversityGate embeds each accepted row and drops near-duplicates above similarity_threshold. The default embedding model is all-MiniLM-L6-v2. Coverage can be measured on an optional metadata field. | In this context, 'embedding model' refers to the specific type of deep learning model used for representing and processing textual data into numerical vectors that can be used for tasks such as text classification or sentiment analysis. |
What does the `probe_temperatures` variable represent? | DiagnosticProbe re-generates a rejected row at several temperatures. probe_temperatures is the list of values. probe_score_split decides which retries count as recovered. Recovered rows return to passed with provenance. | The `probe_temperatures` variable is used to specify the range or temperature for diagnostic probes. |
What is the purpose of enabling 'reward_refiner'? | RewardRefiner rewrites a low-reward answer and re-runs RewardGate. enable_reward_refiner attaches after the reward step. Recovered rows are merged into passed only after the pipeline returns so exporters see them. | Enabling 'reward_refiner' allows for the re-writing of a low-reward answer and re-running it within the same reward pipeline. This feature enables the exporter to maintain or recover data that was previously filtered out due to the low reward. |
What are the main features of SecretGate? | SecretsGate scans instruction and output for API keys, tokens, and high-entropy strings. detect-secrets plugins cover AWS, GitHub, Slack, and generic entropy. Hits become RejectedSample rows, not silent drops. | SecretGate's main features include scanning instructions and outputs for API keys, tokens, and high-entropy strings. It also includes detection plugins for AWS, GitHub, Slack, and generic entropy. |
What does PIIPseudonymizer do? | PIIPseudonymizer runs NER then Faker. The same real name maps to the same alias for the whole run when pii_faker_seed is fixed. Emails, phones, and locations are replaced in the configured fields only. | PIIPseudonymizer runs two main functionalities: NER (Named Entity Recognition) and Faker (Faking Real Names). |
What is ToxicityGate and how does it work? | ToxicityGate runs Detoxify first. Scores below the pass threshold keep the row. Scores above the reject threshold drop it. The band in between can go to an optional LLM judge. | ToxicityGate is a platform that provides a way for individuals to identify and remove toxic substances from their bodies through a series of steps: |
": "Wh | Exact dedup hashes normalized text and keeps the first copy. MinHash estimates Jaccard overlap on character n-grams. embedding_dedup writes a persistent index so later runs still catch repeats. | t is the purpose of exact dedup hashing?",
"answer": "The purpose of exact dedup hashing is to normalize text while keeping only the first copy, which helps in identifying duplicate words or phrases."
}
{
"question": "How does MinHash estimate Jaccard overlap on character n-grams?",
"answer": "MinHash uses chara... |
What does 'clean_fields' function in the context of text processing refer to? | Text cleaning strips HTML, normalizes Unicode, fixes mojibake, collapses whitespace, and drops control characters. clean_fields defaults to instruction, input, and output. Cleaning runs before generation so the LLM never sees the raw junk. | The 'clean_fields' function is used to normalize and clean up text data before it is fed into a language model for generation or understanding. |
How does 'cleaning' work in the process of text generation? | Text cleaning strips HTML, normalizes Unicode, fixes mojibake, collapses whitespace, and drops control characters. clean_fields defaults to instruction, input, and output. Cleaning runs before generation so the LLM never sees the raw junk. | In the context of text generation, 'cleaning' refers to the steps taken to preprocess the input text so that it can be accurately understood and generated by a large language model like BERT. |
What does SchemaGate check for each task type? | SchemaGate checks required fields and token bounds per task type. Pretrain rows need non-empty output. Instruction rows need instruction and output. Failures are RejectedSample records with a structured reason. | SchemaGate checks required fields and token bounds per task type. |
What does the 'exporter' field in the 'output_dir' indicate? | Exporters write Alpaca, ShareGPT, DPO, GRPO, PPO, and corpus JSONL into output_dir. Sidecars are manifest.json, dataset_card.md, and checksums.txt. Empty export files no longer crash a Hub push. | The 'exporter' field in the 'output_dir' represents the name of the software or tool used for exporting data to a specific directory. |
Why is it important to include Manifest.json in the exported files? | Exporters write Alpaca, ShareGPT, DPO, GRPO, PPO, and corpus JSONL into output_dir. Sidecars are manifest.json, dataset_card.md, and checksums.txt. Empty export files no longer crash a Hub push. | Including Manifest.json ensures that the exported files are correctly identified as being part of the same dataset, which is crucial for proper data sharing and integration with other systems. |
What does push_to_hub create? | push_to_hub creates a dataset repo if it is missing and uploads every export plus sidecars. push_format_to_hub sends one format. push_rejected_to_hub publishes the rejected split with the same provenance files. | push_to_hub creates a dataset repo if it is missing. |
What happens when push_to_hub uploads every export plus sidecar? | push_to_hub creates a dataset repo if it is missing and uploads every export plus sidecars. push_format_to_hub sends one format. push_rejected_to_hub publishes the rejected split with the same provenance files. | Every exported file plus sidecar are uploaded to push_to_hub. |
What does max_samples do? | max_samples truncates the passed list after gates. Per-reader max_samples on a dataset dict caps that source before concat. Use the dict form when several Hub sets are mixed. | max_samples truncates the passed list after gates. It specifies how many elements to remove from each batch during data loading. |
How does per-reader max_samples work? | max_samples truncates the passed list after gates. Per-reader max_samples on a dataset dict caps that source before concat. Use the dict form when several Hub sets are mixed. | Per-reader max_samples is used for datasets where multiple Hub sets are mixed together. This ensures that each Hub set gets a fair share of the data. |
What does the 'checkpoint' option in the checkpoint dumping process refer to? | Checkpoints dump generation batches under output_dir/.checkpoints. Re-running the same config resumes instead of repeating completed LLM calls. Disable with enable_checkpoint=False for short Colab smoke tests. | The 'checkpoint' option refers to the directory where checkpoints are saved. It is used to resume or restart the training process after completing a checkpointed model. |
What are LiteLLM and its main functionalities? | LiteLLM is the default generation backend. llm_concurrency is the async worker pool. llm_timeout and llm_max_retries apply per call. Custom api_base points at vLLM or Ollama without changing the rest of the config. | LiteLLM is the default generation backend that provides asynchronous workers to handle multiple tasks concurrently. It also includes llm_concurrency for managing concurrency within each worker. The llm_timeout parameter controls the maximum time allowed between calls to the API. Additionally, llm_max_retries specifies ... |
What does custom api_base point to when using LiteLLM with VLLM or Ollama? | LiteLLM is the default generation backend. llm_concurrency is the async worker pool. llm_timeout and llm_max_retries apply per call. Custom api_base points at vLLM or Ollama without changing the rest of the config. | Custom api_base in LiteLLM refers to the specific implementation used by the backend service. When using LiteLLM with VLLM or Ollama, you specify which version of the service to use by setting this parameter. This allows you to choose between the VLLM and Ollama versions based on your requirements. |
|
curatorkit-testrun-Hallucination
Built using CuratorKIT — provenance-grounded curation and synthesis for LLM post-training.
| Method | qa |
| Backend | litellm |
| Model | openai/Qwen/Qwen2.5-0.5B-Instruct |
| Formats | alpaca |
| Artifact | dataset |
| Published | 2026-08-28 10:47 UTC |
Usage
from datasets import load_dataset
ds = load_dataset("ram-lexsi/curatorkit-testrun-Hallucination", "alpaca")
- Downloads last month
- 27