eval-bfcl-0.5b.py

#!/usr/bin/env python3
"""
BFCL-style eval for the SakThai 0.5B on the held-out combined-v6 `test` split (113).

Measures the numbers that actually matter for this model:
  • simple       — exactly 1 gold tool call  -> did it call the right function?
  • parallel     — >1 gold tool calls         -> did it cover all of them?
  • irrelevance  — 0 gold tool calls (direct)  -> did it correctly NOT call a tool?

This is a pragmatic BFCL-STYLE harness (function-name / abstention scoring), not the
official Berkeley leaderboard. Run it on the OLD repo and the NEW one to get a
before/after. Free — runs on a T4 (or CPU for the 0.5B, just slower).

  MODEL_ID=Nanthasit/sakthai-context-0.5b-merged python eval_bfcl_0.5b.py
"""
import os, re, json, collections
import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = os.environ.get("MODEL_ID", "Nanthasit/sakthai-context-0.5b-merged")
MAX_NEW  = 256
print(f"Model: {MODEL_ID}")

tok = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, torch_dtype=torch.bfloat16,
    device_map="auto" if torch.cuda.is_available() else None,
)
model.eval()

test = load_dataset("Nanthasit/sakthai-combined-v6", split="test")

_TC = re.compile(r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.DOTALL)

def gold_calls(msg):
    """Function names the gold assistant turn calls (OpenAI tool_calls or inline)."""
    names = []
    for tc in (msg.get("tool_calls") or []):
        fn = (tc.get("function") or {}).get("name")
        if fn:
            names.append(fn)
    if not names:  # some rows embed the call in content
        for m in _TC.findall(msg.get("content") or ""):
            try:
                names.append(json.loads(m).get("name"))
            except Exception:
                pass
    return [n for n in names if n]

def pred_calls(text):
    """Function names the model emitted in its generated text."""
    names = []
    for m in _TC.findall(text):
        try:
            names.append(json.loads(m).get("name"))
        except Exception:
            pass
    return [n for n in names if n]

def first_assistant_idx(msgs):
    for i, m in enumerate(msgs):
        if m.get("role") == "assistant":
            return i
    return None

buckets = collections.defaultdict(lambda: [0, 0])  # category -> [correct, total]

for ex in test:
    msgs = ex["messages"]
    tools = ex.get("tools") or None
    idx = first_assistant_idx(msgs)
    if idx is None or idx == 0:
        continue
    prompt_msgs, gold = msgs[:idx], msgs[idx]
    gold_names = gold_calls(gold)
    cat = "irrelevance" if not gold_names else ("simple" if len(gold_names) == 1 else "parallel")

    inputs = tok.apply_chat_template(
        prompt_msgs, tools=tools, add_generation_prompt=True, return_tensors="pt",
    ).to(model.device)
    with torch.no_grad():
        out = model.generate(inputs, max_new_tokens=MAX_NEW, do_sample=False,
                             pad_token_id=tok.eos_token_id)
    gen = tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True)
    pred_names = pred_calls(gen)

    if cat == "irrelevance":
        ok = len(pred_names) == 0                       # correct = abstained
    elif cat == "simple":
        ok = gold_names[0] in pred_names                # right function called
    else:  # parallel
        ok = set(gold_names).issubset(set(pred_names))  # covered all gold calls
    buckets[cat][0] += int(ok)
    buckets[cat][1] += 1

print(f"\nBFCL-style results — {MODEL_ID}")
print(f"{'category':<14}{'pass':>6}{'total':>7}{'acc':>8}")
tc = tt = 0
for cat in ("simple", "parallel", "irrelevance"):
    c, t = buckets[cat]
    tc += c; tt += t
    acc = f"{100*c/t:5.1f}%" if t else "  n/a"
    print(f"{cat:<14}{c:>6}{t:>7}{acc:>8}")
print(f"{'OVERALL':<14}{tc:>6}{tt:>7}{(f'{100*tc/tt:5.1f}%' if tt else '  n/a'):>8}")
print("\nNote: BFCL-STYLE (function-name / abstention scoring), not official BFCL.")