DataForge SFT Trajectories: Supervised Fine-Tuning Records for AI Model Warmup

A dataset containing chunk-level supervised-fine-tuning records for the DataForge warmup model, including expert versions v1 through v4. The milestone data is built from split-safe CSV diffs, anchoring training to audited labels. It was created by Praneshrajan15 and last updated on 2026-05-19.

Use Cases

• Training AI models using supervised fine-tuning based on the described expert-v1 to expert-v4 records.
• Auditing model training processes based on the use of audited labels rather than teacher guesses.
• Comparing model performance across different training checkpoints, such as the v0-smoke checkpoint.
• Studying the impact of contract-repair and inferability-auditing steps on model training trajectories.

Strengths

• Training is anchored to audited labels rather than teacher guesses, as stated in the description.
• Includes multiple expert versions (v1 through v4) of supervised-fine-tuning records.
• The dataset page indicates a last update timestamp of 2026-05-19 15:16:38.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Built from split-safe dirty/clean CSV diffs (oracle_from_clean_diff).

Freshness

Last updated 2026-05-19 15:16:38; freshness should be verified.