README.md

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- sft
- synthetic
- instruction-tuning
- function-calling
- reasoning
- code
pretty_name: Unified SFT Dataset
size_categories:
- 100K<n<1M
---

# Unified SFT Dataset

A supervised fine-tuning (SFT) mixture built for a 16,384-token context
model, combining chat quality, general knowledge, agentic tool use,
reasoning, code, strict instruction-following, rewriting/summarization,
and system-prompt-adherence data into a single shuffled `train` split.

Every row was kept only if its full conversation, rendered through the
tokenizer's chat template, fit inside **16,000 tokens** (small headroom
under a 16,384 hard limit).

## Source datasets

| Source | Upstream dataset | Purpose |
|---|---|---|
| Smol-Magpie-Ultra | `HuggingFaceTB/smoltalk` (smol-magpie-ultra) | Chat depth & quality |
| OpenHermes-2.5 | `HuggingFaceTB/smoltalk` (openhermes-100k) | Generalization / MMLU-style |
| APIGen-Function-Calling | `HuggingFaceTB/smoltalk` (apigen-80k) | Agentic tool use |
| Orca-AgentInstruct-1M | `microsoft/orca-agentinstruct-1M-v1` (analytical_reasoning, fermi, brain_teaser) | Reasoning |
| Self-OSS-Starcoder2-Instruct | `HuggingFaceTB/smoltalk` (self-oss-instruct) | Code agents |
| Smol-Constraints | `HuggingFaceTB/smoltalk` (smol-constraints) | Strict IFEval-style compliance |
| Smol-Rewrite / Smol-Summarize | `HuggingFaceTB/smoltalk` (smol-rewrite, smol-summarize) | Extraction & rewriting |
| SystemChats-2.0 | `HuggingFaceTB/smoltalk` (systemchats-30k) | System prompt adherence |

## Composition (target mix)

| Source | Rows (target) | % |
|---|---:|---:|
| Smol-Magpie-Ultra | 350,000 | 43.75% |
| OpenHermes-2.5 | 100,000 | 12.50% |
| APIGen-Function-Calling | 80,000 | 10.00% |
| Orca-AgentInstruct-1M | 100,000 | 12.50% |
| Self-OSS-Starcoder2-Instruct | 50,000 | 6.25% |
| Smol-Constraints | 36,000 | 4.50% |
| Smol-Rewrite & Smol-Summarize | 50,000 | 6.25% |
| SystemChats-2.0 | 34,000 | 4.25% |
| **Total** | **800,000** | **100%** |

These are *targets*, not guarantees — each source is streamed and
filtered live, so the actual kept count can come in lower if a source
runs out of in-budget rows before its target is reached (the script
logs a warning and reports the real final counts at the end of the run).

## Data fields

| Field | Type | Description |
|---|---|---|
| `source` | string | Group label (e.g. `"Orca-AgentInstruct-1M"`) |
| `subset` | string | Specific config/split (e.g. `"Orca-AgentInstruct-1M (fermi)"`) |
| `messages` | list[{role, content}] | The conversation |
| `num_tokens` | int | Token count of the rendered conversation (chat-template applied) |

## Data splits

Single `train` split. No validation/test split is carved out.

## Construction details

- Tokenizer: `Qwen/Qwen3.5-0.8B` chat template (swap for your target model's tokenizer if different)
- Max context: 16,000 tokens per conversation
- Each source shuffled (buffer 10,000) before sampling, seed 42
- Rows re-shuffled globally across sources before saving
- A safety cap stops scanning a source after `target × 15` rows if the target can't be filled

## Licensing

This mixture inherits the license of each upstream source rather than
carrying one license of its own — `HuggingFaceTB/smoltalk`'s component
configs and Microsoft's `orca-agentinstruct-1M-v1` each have their own
terms. Check the upstream dataset pages before redistributing or using
this mixture commercially.

## Citation

Please cite the original upstream datasets (smoltalk / SmolLM team,
OpenHermes, APIGen, Orca-AgentInstruct-1M, Self-OSS-Starcoder2-Instruct)
rather than this repo alone, since all the underlying data originates there.