| """Tool-call calibration conversations: NousResearch/hermes-function-calling-v1 (func-calling.json, |
| rev dae3e1d), parsed into structured tool_calls and rendered through the model's own chat template |
| with tools=, so the Hessians see the served <tools> system prompt and <tool_call><function=...> turns. |
| No BFCL data. Import render(tok, path) -> list of rendered strings, in file order.""" |
| import json, re |
|
|
| CALL = re.compile(r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.S) |
| RESP = re.compile(r"<tool_response>\s*(.*?)\s*</tool_response>", re.S) |
|
|
|
|
| def convert(r): |
| tools = json.loads(r["tools"]) if isinstance(r["tools"], str) else r["tools"] |
| msgs = [] |
| for m in r["conversations"]: |
| v = m["value"] |
| if m["from"] == "system": |
| continue |
| if m["from"] == "human": |
| msgs.append({"role": "user", "content": v}) |
| elif m["from"] == "gpt": |
| calls = [json.loads(c) for c in CALL.findall(v)] |
| if calls: |
| msgs.append({"role": "assistant", "content": CALL.sub("", v).strip(), |
| "tool_calls": [{"type": "function", "function": {"name": c["name"], "arguments": c.get("arguments", {})}} for c in calls]}) |
| else: |
| msgs.append({"role": "assistant", "content": v}) |
| elif m["from"] == "tool": |
| for body in RESP.findall(v) or [v]: |
| msgs.append({"role": "tool", "content": body}) |
| return tools, msgs |
|
|
|
|
| def render(tok, path, **kw): |
| out = [] |
| for r in json.load(open(path)): |
| try: |
| tools, msgs = convert(r) |
| out.append(tok.apply_chat_template(msgs, tools=tools, tokenize=False, **kw)) |
| except (ValueError, KeyError, TypeError, json.JSONDecodeError): |
| continue |
| return out |
|
|
