---
license: apache-2.0
language:
- en
- fr
task_categories:
- text-generation
- conversational
size_categories:
- 100K<n<1M
tags:
- chatml
- sft
- aether-family
- code
configs:
- config_name: default
  data_files:
  - split: train
    path: "code_shard_*.jsonl"
---

# archon-sft-v1-code

AETHER family SFT dataset — group `code`.

**Format**: JSONL ChatML messages, task_type tagged, MinHash dedup applied (threshold 0.85).

**Schema**:
```
{
  "messages": [{"role": "system|user|assistant", "content": "..."}],
  "task_type": "function_calling|code|reasoning_cot|...",
  "source_ds": "<HF dataset_id>",
  "lang": "en|fr|...",
  "system_source": "archon_default|overridden_from_source"
}
```

Generated by `prepare_sft.py` pipeline (2026-05-25).

<!-- AUDIT_2026-05-28 -->

## Real sources sampled 2026-05-28

- 85.7% `ise-uiuc/Magicoder-OSS-Instruct-75K`
- 14.3% `m-a-p/CodeFeedback-Filtered-Instruction`

### Real task_type distribution (sampled)

- 51% `code`
- 48% `math_symbolic` (auto-tagged from code with formulas)
- 1.4% `protobuf_grpc`

### Language reality

Sampled 700 rows across multiple offsets: **100% English** detected despite `language: [en, fr]` tag. The `lang` field is consistently `en`.

### Notes

Code instruction-following SFT.
This is the ARCHON SFT v1 reality, not the planning document. Memory/plan files may list different intended sources — those represent the initial design, not what was actually staged.