nemotron-code-oracle-qwen3.5-122b-opencode-sft

# nemotron-code-oracle · Qwen3.5-122B-A10B · opencode · SFT Multi-turn agentic SFT conversations distilled from opencode coding-agent traces generated with **Qwen/Qwen3.5-122B-A10B-FP8** (131k context, verifier-on). Each **assistant** turn is decoded **verbatim** from the serving engine's literal completion token IDs (RecordProxy capture) with the exact `Qwen/Qwen3.5-122B-A10B-FP8` tokenizer — so it carries the model's real reasoning (`<think>…</think>`) and native opencode tool calls (`<tool_call><function=…><parameter=…>`). System + task turns are decoded from the first-step prompt; tool/observation turns come from the trace. Rows without literal tokens are dropped. ## Fields - `conversations` — ShareGPT `[{from: system|human|gpt, value}]`. - `text` — the flat conversation with Qwen `<|im_start|>` markers, **reasoning preserved on every turn** (the stock chat template strips `<think>` from history; this string does not, so per-turn reasoning is available for SFT). - `task`, `num_turns`. Source traces: `penfever/nemotron-code-oracle-filtered-qwen3.5-122b-131k-opencode-traces`. Converter: literal completion tokens → text via `AutoTokenizer.from_pretrained("Qwen/Qwen3.5-122B-A10B-FP8")`.