OsaurusAI/raptor-0.5-harness-corpus

Raptor 0.5 harness corpus

The SFT corpus used to produce Raptor 0.5 from inclusionAI/Ling-3.0-tiny. It is small on purpose: the goal was tool-surface familiarity for the Osaurus harness, not new capabilities.

Contents

1,150 rows total, 230,034 assistant-supervised tokens, max sequence 4,096. Trained as a rank-4 LoRA (alpha 8) and merged into base weights at BF16 after 4 optimizer updates.

Row format

Rows are pre-rendered against the Ling-3.0-tiny chat template (revision b61f4338, template SHA-256 recorded per row) and carry both the readable and the tokenized form:

• —messages, tools — the conversation and the tool menu offered (may be empty)
• —native_input_ids, native_labels — tokenized sequence; labels are -100 outside supervision_spans (tool observations are always masked)
• —availability — tools_absent / relevant menu / irrelevant menu
• —system_condition — which system-prompt variant the row was rendered under
• —reasoning_mode — on / off, bound per row
• —executable_verification — for rows grounded in runnable projects: language, run command, expected-stdout hash, and PASS_EXECUTABLE status
• —provenance hashes for the rendered text, source profile, tool menu, and system prompt, plus the source commit of the generator

Composition

• —reasoning on / off: 575 / 575
• —tool availability: 168 tools-absent · 148 relevant menu · 32 irrelevant menu
• —system condition: 144 current harness prompt · 108 no system · 96 generic non-Osaurus prompt
• —expected actions per row: 200 zero-action · 18 single-call · 130 multi-call

Zero-action rows are deliberate: answering directly when no offered tool fits is trained, not assumed.

How rows were admitted

Every candidate row passed a mechanical gate before it entered the corpus: render through the real chat template, every tool call parses, every called tool exists in the row's own tools[], every argument validates against the real schema, reasoning tags structurally correct for the row's mode, tokenizer round-trip lossless on special tokens, near-duplicate scan, and a plain-text final answer present. Tool schemas and system prompts were extracted from the Osaurus host source, not written by hand.

Notes

• —The tool-call dialect is Ling's native XML-arg format (<arg_key> / <arg_value> inside <tool_call>), while the tool definitions the model reads are JSON. Parsers that expect JSON tool calls will not match.
• —The corpus was verified do-no-harm: the merged model's failure set on a 144-run defect sweep is identical to the base model's, and base benchmark scores are unchanged.

Osaurus · eric@osaurus.ai