rajnadkarni/sera-django-alexaprize-sft

Sera Django AlexaPrize Tool-Call SFT

This dataset packages converted allenai/Sera-4.5A-Django-T2 trajectories in the AlexaPrize coding-agent tool-call format. It is intended for supervised fine-tuning of a coding agent that uses Bedrock Converse-style messages and tool calls.

Contents

• —train: 24,250 records
• —validation: 494 records
• —Synthetic update_pr examples in train: 3,000
• —Format marker: bedrock_converse_v1
• —Tool schema: tool_schemas/alexaprize_tools.json

Each row has:

• —id: stable record id
• —format: bedrock_converse_v1
• —system: coding-agent system prompt
• —messages: ordered user/assistant/tool-use/tool-result conversation
• —metadata: source, repository, branch, expected action, and synthetic-update fields

Tool Format

The records preserve Bedrock Converse-style content blocks:

• —assistant tool calls use {"toolUse": {"toolUseId", "name", "input"}}
• —tool outputs use {"toolResult": {"toolUseId", "content", "status"}}
• —text blocks use {"text": "..."}

The companion schema file lists the AlexaPrize tool definitions used for validation.

Provenance

Derived from allenai/Sera-4.5A-Django-T2, downloaded from the default branch on 2026-05-15, and extended with synthetic update_pr examples generated by NVIDIA NeMo Data Designer against a local Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8 vLLM endpoint.

Source license is declared as ODC-By-v1.0 by the source dataset. Preserve attribution to the original dataset and include this card/manifest when handing off the data.

Validation

The release was validated against tool_schemas/alexaprize_tools.json:

• —train records validated: 24,250
• —validation records validated: 494
• —invalid records: 0
• —train create_pr records: 24,250
• —train update_pr records: 3,000

Detailed counts and SHA-256 hashes are in manifest.json.