Datasets:
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Error code: UnexpectedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
id string | source_file_id string | source_repo_url string | messages list | task_type string | source_language null | model_used string | license string | schema_version int64 |
|---|---|---|---|---|---|---|---|---|
sft_gleam_00001 | 0918nobita/domu/src/domu/console.gleam | https://github.com/0918nobita/domu | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00002 | 0918nobita/domu/src/domu/global/console.gleam | https://github.com/0918nobita/domu | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00003 | 0918nobita/domu/src/domu/internal/iterator_compat.gleam | https://github.com/0918nobita/domu | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00004 | 0918nobita/domu/src/domu/window.gleam | https://github.com/0918nobita/domu | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00005 | 0918nobita/domu/test/domu_test/event_target.gleam | https://github.com/0918nobita/domu | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00006 | 0918nobita/domu/test/happy_dom.gleam | https://github.com/0918nobita/domu | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00007 | 0918nobita/json_canvas/src/json_canvas/internal/decode/edge.gleam | https://github.com/0918nobita/json_canvas | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00008 | 0918nobita/json_canvas/src/json_canvas/internal/encode/edge.gleam | https://github.com/0918nobita/json_canvas | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00009 | 0918nobita/json_canvas/src/json_canvas/internal/generic_node.gleam | https://github.com/0918nobita/json_canvas | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00010 | 0ngk/kitazith/src/kitazith/component.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00011 | 0ngk/kitazith/src/kitazith/component/separator.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00012 | 0ngk/kitazith/src/kitazith/flag.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00013 | 0ngk/kitazith/src/kitazith/webhook/delete_query.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00014 | 0ngk/kitazith/test/kitazith/message_formatting/guild_navigation_test.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00015 | 0ngk/kitazith/test/kitazith/message_formatting/timestamp_test.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00016 | 0ngk/kitazith/test/kitazith/poll_test.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00017 | 0ngk/kitazith/test/kitazith/timestamp_test.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00018 | 0ngk/kitazith/test/kitazith/webhook/delete_query_test.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00019 | 0ngk/kitazith/test/kitazith/webhook/edit_query_test.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
sft_gleam_00020 | 0ngk/kitazith/test/kitazith/webhook/get_query_test.gleam | https://github.com/0ngk/kitazith | [
{
"role": "system",
"content": "You are a helpful Gleam programming assistant.\n\nKey Gleam rules:\n- Float operators use a dot: +. -. *. /. String concatenation uses <>.\n- No return keyword — the last expression is returned. No null or exceptions.\n- Error handling uses Result(success, error) with Ok/Err... | code_gen | null | gpt-oss-120b | Apache-2.0 | 1 |
Gleam Instruct
A synthetic instruction-following dataset for fine-tuning language models to become Gleam programming specialists. Contains 14,053 high-quality instruction-response pairs covering code generation, explanation, and completion tasks, all derived from real open-source Gleam code.
Dataset Summary
| Metric | Value |
|---|---|
| Total records | 14,053 |
| Task types | code_gen, explanation, completion |
| Source files | 4,700 unique .gleam files |
| Source repos | 997 GitHub repositories |
| Format | ChatML (system → user → assistant) |
| Generation model | gpt-oss-120b (primary), openai/gpt-oss-120b via OpenRouter, glm-5.1 via Z.ai |
By task type
| Task | Records | Description |
|---|---|---|
code_gen |
4,667 | Write a new Gleam module inspired by a reference snippet |
explanation |
4,698 | Explain a Gleam code snippet in detail |
completion |
4,688 | Complete a partially-truncated Gleam module |
Schema
Each record follows standard ChatML format with metadata:
{
"id": "sft_gleam_00001",
"source_file_id": "owner/repo/path.gleam",
"source_repo_url": "https://github.com/owner/repo",
"messages": [
{"role": "system", "content": "You are a helpful Gleam programming assistant..."},
{"role": "user", "content": "<task-specific instruction with source code>"},
{"role": "assistant", "content": "<model response>"}
],
"task_type": "code_gen",
"source_language": null,
"model_used": "gpt-oss-120b",
"license": "Apache-2.0",
"schema_version": 1
}
Field descriptions
| Field | Type | Description |
|---|---|---|
id |
string | Unique record identifier |
source_file_id |
string | Source file in owner/repo/path.gleam format |
source_repo_url |
string | GitHub repository URL for attribution |
messages |
list | ChatML conversation (system + user + assistant) |
task_type |
string | code_gen, explanation, or completion |
source_language |
null | Reserved for translation tasks (future) |
model_used |
string | Model that generated the response |
license |
string | Dataset license |
schema_version |
int | Schema version (currently 1) |
Task Type Details
Code Generation (code_gen)
Given a Gleam code snippet as a reference, write a new module that solves a related but different problem. The model is instructed to use idiomatic Gleam: type annotations, pattern matching, Result for errors, and pipe operators.
Explanation (explanation)
Given a Gleam code snippet, provide a detailed explanation covering what each function does, how types are used, and notable patterns or idioms. Explanations are targeted at someone learning Gleam.
Completion (completion)
Given the first ~60% of a Gleam module (truncated at a random line with a // TODO hint), complete the remaining functions to produce a working module with type annotations and idiomatic patterns.
Generation Pipeline
- Corpus: 22,581 Gleam source files from 1,528 GitHub repositories (see kasuboski/gleam-code-corpus)
- Sampling: 4,700 files sampled with seed=42, filtered to exclude oversized files (>3K input tokens)
- Generation: Each file used to generate one instruction-response pair per task type via
openai/gpt-oss-120b - Deduplication: Kept latest generation per (file, task_type) — some files were regenerated to fix truncation
- Filtering: Removed truncated responses, failed API calls, and 6 degenerate repetitive outputs (unique line ratio < 0.3)
- Reconstruction: User messages rebuilt from source code + task templates; assistant messages preserved verbatim
Quality
- 0% truncation rate on explanation tasks (max_tokens bumped to 8,192)
- <1% truncation on code_gen (0.6%) and completion (0.2%)
- 6 degenerate records filtered out (model looping on repeated patterns)
- All source code from permissive-license repositories
Loading the Dataset
from datasets import load_dataset
ds = load_dataset("kasuboski/gleam-instruct", data_files="train.jsonl")
Or directly with polars:
import polars as pl
df = pl.read_ndjson("gleam-instruct.jsonl")
Training-Time Mixing
This dataset contains only Gleam-focused pairs. For best results during SFT, we recommend mixing with anti-forgetting data at training time:
| Source | ~Count | Purpose |
|---|---|---|
| This dataset | 14,053 | Gleam specialization |
| HuggingFaceH4/ultrachat_200k | ~3,000 | General instruction following |
| nvidia/OpenCodeInstruct | ~500 | General code skill preservation |
A ~30% general data mix prevents catastrophic forgetting when fine-tuning on a narrow domain like Gleam. Adjust ratios based on your base model and use case.
Intended Use
- Supervised fine-tuning of LLMs for Gleam code generation and understanding
- Domain adaptation experiments for niche programming languages
- Instruction tuning research on code-specific tasks
Limitations
- Responses are synthetically generated and may contain subtle errors or non-idiomatic patterns
- Code generation tasks use a limited set of 5 instruction templates
- No syntax validation was performed on generated Gleam code
- The dataset is English-only
- Source code is weighted toward popular/maintained repos — may underrepresent edge cases
License
Apache-2.0
- Downloads last month
- 21