CoolFace
Datasetpublic

JacktheLander/fungi-rag-agent-sft-database

Fungi RAG Agent SFT Database This dataset contains the supervised fine-tuning database used to train the JacktheLander/smollm2-1.7b-fungi-rag-agent-distill-lora-gguf adapter. It was built for the project-owned fungi RAG learning system: JacktheLander/FunghiResearchAgent The examples teach a small local language model to follow the system's agent contract: use rag.search for evidence-backed mycology answers, use safety.review for wild-mushroom edibility, field-identification… See the full description on the dataset page: https://huggingface.co/datasets/JacktheLander/fungi-rag-agent-sft-database.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
0likes15downloads
Dataset Card

Fungi RAG Agent SFT Database

This dataset contains the supervised fine-tuning database used to train the JacktheLander/smollm2-1.7b-fungi-rag-agent-distill-lora-gguf adapter.

It was built for the project-owned fungi RAG learning system:

`JacktheLander/FunghiResearchAgent`

The examples teach a small local language model to follow the system's agent contract:

  • —use rag.search for evidence-backed mycology answers,
  • —use safety.review for wild-mushroom edibility, field-identification, and other safety-sensitive prompts,
  • —cite only numeric source IDs returned by retrieval tools,
  • —avoid unsupported claims,
  • —refuse or redirect unsafe identification/edibility requests,
  • —express uncertainty when evidence is partial, missing, or conflicting.

Splits

The exact prepared files used for SFT are in prepared/:

FileExamplesDescription
prepared/train.jsonl3,387Training split rendered for chat-template SFT
prepared/val.jsonl298Validation split
prepared/test.jsonl300Held-out test split

Source export files are in source_exports/:

FileDescription
source_exports/sft_train.jsonlStructured source SFT training records
source_exports/sft_val.jsonlStructured source SFT validation records
source_exports/sft_test.jsonlStructured source SFT test records
source_exports/redteam_safety.jsonlSafety/red-team evaluation prompts

Validation and dataset statistics are in metadata/.

Data Quality Summary

Prepared dataset metrics:

MetricValue
Total prepared examples3,985
Schema valid rate1.000
Average token length449.31
P95 token length662
Max token length946
rag.search tool calls2,445
safety.review tool calls1,051
Examples with citations2,425
Unsupported citation examples0
Citation failure count0
Safety failure count0
Missing required RAG calls0
Duplicate or near-duplicate count after preparation0

Format

Prepared records are JSON Lines. Each row includes fields such as:

  • —example_id
  • —split
  • —text
  • —messages
  • —training_messages
  • —metadata
  • —tools
  • —token_length
  • —assistant_only_loss

The text field is the rendered chat-template training sequence. The structured messages and training_messages fields are included to make future re-rendering and audit possible.

Source And License Notes

This dataset combines generated agent examples, tool-call scaffolding, and retrieved evidence snippets used by the Fungi RAG project. The upstream source material includes open educational and scientific references with varying source licenses. Review the source metadata and original licenses before redistribution or commercial use.

This dataset is released for research and educational evaluation of the linked fungi RAG agent.

Safety

This dataset contains examples about fungi, including safety-sensitive prompts involving wild mushroom identification and edibility. The desired behavior is refusal or expert redirection.

Do not use this dataset or any model trained from it to make mushroom edibility, field-identification, medical, dosage, or emergency toxicology decisions.

Related Model

`JacktheLander/smollm2-1.7b-fungi-rag-agent-distill-lora-gguf`