biscet/DFlash-TraceMix-Lite

DFlash TraceMix Lite

A small dataset for the next draft.

1,996 records  ·  3 stages  ·  JSONL

Explore the files · Manifest · Processing report

---

Built for DFlash experiments. A lightweight collection of prompts and model-generated response traces for training and iterating on draft models.

Project code, programming, math, knowledge and writing — a compact mix for your next training run.

Repository branches

Branch	Contents
main	The current dataset card and repository guide.
v1.0.0	The complete release: JSONL data, manifest, processing report and release card.

Looking for the data? Open the v1.0.0 branch. When downloading through the Hub API, use v1.0.0 as the revision.

The mix

File	Focus	Train	Validation
pilot.jsonl	A small starting batch	8	4
project.jsonl	Source-code understanding	512	64
openmix.jsonl	Mixed tasks + project replay	1,280	128
Total		1,800	196

Each record pairs a prompt with a generated completion, alongside split, task and source metadata.

Before your training run

• Target model: existing traces come from Whittle Q5. For another DFlash target, adapt the chat template, regenerate responses and prepare tokenization and target features for your pipeline. Original token IDs and hidden features are not included. See the training guidance.
• Splits and replay: eval denotes validation. OpenMix reuses 256 Project Code training records; they are not additional unique examples. Four short OpenMix training traces have eligible_for_original_loader: false.
• Publication copy: text was anonymized after training; the original weights were not retrained or unlearned. The edited traces are not verified reference answers, and complete removal of identifying details is not guaranteed. See the processing report.

Attribution & terms

Release license: Apache-2.0. Original contributions and release packaging are provided under this license. Third-party source terms remain applicable; attribution is recorded in source_dataset, source_subset and source_revision.

Project Code was authored and curated by Egor Prokhorov.