datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Hemlock-SFT
Hemlock-SFT
This dataset was built to teach LLMs The Hemlock Programming Language using SFT.
Please note: this dataset is obsolete with the release of Hemlock 2.0.0. Please use: hemlang/Hemlock2-DPO
hemlock-codex3-SFT
hemlock-codex3-SFT
Code-generation SFT data for the Hemlock
programming language: translation tasks (C/Go/JavaScript/Python/Rust → Hemlock),
algorithm/systems generation, and stdlib recall — every reference answer
execution-verified against the Hemlock interpreter at build time.
Supersedes and merges hemlock-codex-SFT
(a strict subset of codex2), hemlock-codex2-SFT,
and hemlock-formulary-SFT.
What changed vs codex2/formulary
Fenced outputs. Every output is a… See the full description on the dataset page: https://huggingface.co/datasets/hemlang/hemlock-codex3-SFT.hemlock-codex-SFT
Hemlock Codex SFT
Supervised fine-tuning dataset for the Hemlock programming language. Contains 552 instruction/output pairs covering algorithms, systems programming, cross-language translation, and practical programs.
Motivation
Benchmark results for hemlang/Hemlock2-Coder-7B (Q8_0, zero-shot) showed weak performance on:
L3 Algorithms (28.6%) — data structures, graph algorithms, DP
L4 Systems Programming (42.9%) — manual memory, concurrency patterns
L5 Translation… See the full description on the dataset page: https://huggingface.co/datasets/hemlang/hemlock-codex-SFT.hemlock-transmutation
hemlock-transmutation
Execution-verified C → Hemlock and Python → Hemlock translation pairs, mined from The Stack v3.
Every row is a real program from GitHub, its byte-exact stdout captured by actually running it,
and a Hemlock translation that was executed and confirmed to reproduce that stdout. Nothing is
included on a model's say-so.
rows (clean)
1,966
rows (all, incl. flagged)
2,423
source languages
C (1,102), Python (864)
source repositories
655… See the full description on the dataset page: https://huggingface.co/datasets/hemlang/hemlock-transmutation.Hemlock-SFT-combined
Hemlock-SFT-combined
The three Hemlock SFT sources merged into one instruction / output set for single-job
fine-tuning, with benchmark-overlapping rows removed.
5,852 rows.
source
rows
what it teaches
hemlang/Hemlock-SFT
2,673
the language itself — docs, concepts, parity tests, examples
hemlang/hemlock-codex3-SFT
1,213
task-shaped generation ("model a bank account using a define type…")
hemlang/hemlock-transmutation
1,966
translating C, Python and others into… See the full description on the dataset page: https://huggingface.co/datasets/hemlang/Hemlock-SFT-combined.egirl-hemlock-dpo
egirl-hemlock-DPO
400 pairs teaching two things at once: write Hemlock, and write it through a tool call.
Each row asks for a C or Python program to be translated to Hemlock
and saved to a path. The target is a single write_file call whose content argument is the
Hemlock translation — so the response shape teaches the tool interface while the argument teaches
the language.
Built for egirl, whose local operator model reaches for
tools rather than answering in prose. Companion to… See the full description on the dataset page: https://huggingface.co/datasets/schneewolflabs/egirl-hemlock-dpo.hemlock-formulary-SFT
Hemlock Apothecary Formulary
Stdlib-focused SFT data for fine-tuning Hemlock-Apothecary-7B — a
Hemlock-Codex-7B derivative
aimed at closing the L2-Stdlib gap measured on hembench.
A formulary in pharmacology is the reference book listing drug compositions and dosages.
Same idea here: one realistic program per (@stdlib/<module>, task) pair showing
exactly how to compose real Hemlock stdlib calls — right imports, right method names,
right idioms.
Why this dataset exists… See the full description on the dataset page: https://huggingface.co/datasets/hemlang/hemlock-formulary-SFT.Hemlock2-DPO
Hemlock2-DPO
A Direct Preference Optimization (DPO) training dataset for the Hemlock programming language (v2.0.0).
Dataset Description
852 prompt/chosen/rejected triples covering Hemlock's syntax, standard library, memory management, concurrency model, FFI, and common patterns. Designed to train language models to produce correct Hemlock code and documentation while avoiding cross-language contamination from Python, JavaScript, and C.… See the full description on the dataset page: https://huggingface.co/datasets/hemlang/Hemlock2-DPO.hemlock-codex2-SFT
hemlock-codex2-SFT
v2 of the Hemlock code-generation SFT set: the original 552 codex rows plus an
interpreter-validated expansion targeting the categories the models fail on
(graphs, dp, trees, sorting, search, memory, concurrency, defer,
practical/data-processing), each with codex-style translation variants from all
five source languages (Python, JavaScript, C, Go, Rust).
552 original codex rows
101 new hard generation examples (every one run through the Hemlock
interpreter… See the full description on the dataset page: https://huggingface.co/datasets/hemlang/hemlock-codex2-SFT.hemlock-sft-v0.1
hemlock-sft-v0.1
This dataset was built to teach LLMs The Hemlock Programming Language using SFT.
Data was generated deterministically from existing documentation, tests, and example code. Then cleaned using Claude Sonnet 4.5.
hemlock-sft-v0.2
hemlock-sft-v0.2
This dataset was built to teach LLMs The Hemlock Programming Language using SFT.
Data was generated deterministically from existing documentation, tests, and example code. Then cleaned using Claude Sonnet 4.5.
This updated version was corrected by Claude Opus 4.5 using Claude Code.