job-0.5b-v7.py

# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "transformers>=4.44", "datasets", "accelerate"]
# ///
"""SakThai 0.5B config-upgrade fine-tune on combined-v7, scored on sakthai-bench-v1.

Changes vs the 2026-07-29 run that died at step 60/252:
  * report_to="none" — trackio's config->parquet export cannot serialise PEFT's
    empty `rank_pattern` struct and killed that run at the first checkpoint push
    (pyarrow ArrowNotImplementedError). No logger is worth losing a 2h run.
  * save_strategy="no" + a single push after training — no mid-run hub pushes,
    so an upload hiccup can never destroy a nearly-finished run.
  * trains on combined-v7 minus every row reserved by sakthai-bench-v1 and every
    row that so much as *offers* a held-out tool.
  * eval reads the balanced bench (simple / parallel / irrelevance_tools /
    irrelevance_no_tools) and reports the unseen-tool slice separately.
"""
import re, json, gc, hashlib, collections, urllib.request
import torch
from datasets import load_dataset, concatenate_datasets
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import LoraConfig, get_peft_model, PeftModel
from trl import SFTTrainer, SFTConfig

USER, BASE_MODEL = "Nanthasit", "Qwen/Qwen2.5-0.5B-Instruct"
OLD_MERGED   = f"{USER}/sakthai-context-0.5b-merged"
ADAPTER_REPO = f"{USER}/sakthai-context-0.5b-tools-v2"
MERGED_REPO  = f"{USER}/sakthai-context-0.5b-merged-v2"
DATASET      = f"{USER}/sakthai-combined-v7"
BENCH        = f"{USER}/sakthai-bench-v1"
EXCLUDE_URL  = f"https://huggingface.co/datasets/{BENCH}/resolve/main/train_exclude_fingerprints.json"
MAX_LEN = 1536

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# ── Manual ChatML renderer. Qwen's built-in template cannot render this data:
#    content=None on tool turns, arguments as a JSON string, tool results
#    sometimes lists. Validated locally on real rows before every run. ──────
def _text(c):
    return "" if c is None else (c if isinstance(c, str) else json.dumps(c, ensure_ascii=False))
def _tools_block(tools):
    if not tools: return ""
    sigs = "\n".join(json.dumps(t, ensure_ascii=False) for t in tools)
    return ("\n\n# Tools\n\nYou may call one or more functions. Signatures are within "
            "<tools></tools>:\n<tools>\n" + sigs + "\n</tools>\n\nFor each call return:\n"
            "<tool_call>\n{\"name\": <name>, \"arguments\": <json>}\n</tool_call>")
def _assistant_body(m):
    body = _text(m.get("content"))
    for tc in (m.get("tool_calls") or []):
        fn = tc.get("function", tc); a = fn.get("arguments", "{}")
        if not isinstance(a, str): a = json.dumps(a, ensure_ascii=False)
        body += ("\n" if body else "") + '<tool_call>\n{"name": "%s", "arguments": %s}\n</tool_call>' % (fn.get("name", ""), a)
    return body
def _render_msg(m, tools_sys):
    r = m.get("role")
    if r == "system": return "<|im_start|>system\n" + _text(m.get("content")) + _tools_block(tools_sys) + "<|im_end|>\n"
    if r == "user":   return "<|im_start|>user\n" + _text(m.get("content")) + "<|im_end|>\n"
    if r == "tool":   return "<|im_start|>user\n<tool_response>\n" + _text(m.get("content")) + "\n</tool_response><|im_end|>\n"
    if r == "assistant": return "<|im_start|>assistant\n" + _assistant_body(m) + "<|im_end|>\n"
    return ""
def render_chatml(messages, tools, add_generation_prompt=False):
    messages = messages or []
    out = []
    if not (messages and messages[0].get("role") == "system") and tools:
        out.append("<|im_start|>system\nYou are a helpful assistant." + _tools_block(tools) + "<|im_end|>\n")
    for i, m in enumerate(messages):
        out.append(_render_msg(m, tools if (i == 0 and m.get("role") == "system") else None))
    if add_generation_prompt: out.append("<|im_start|>assistant\n")
    return "".join(out)

def fingerprint(messages):
    return hashlib.sha1(json.dumps(messages, sort_keys=True, ensure_ascii=False).encode()).hexdigest()

# ── Data: v7 minus everything the benchmark reserves ─────────────────────
with urllib.request.urlopen(EXCLUDE_URL) as r:
    _ex = json.load(r)
EXCLUDE = set(_ex["fingerprints"])
HELD_OUT_TOOLS = set(_ex["held_out_tools"])
print(f"bench reserves {len(EXCLUDE)} rows; held-out tools: {sorted(HELD_OUT_TOOLS)}")

def _keep(ex):
    if fingerprint(ex["messages"]) in EXCLUDE:
        return False
    names = {(t.get("function") or {}).get("name") or t.get("name") for t in (ex.get("tools") or [])}
    return not (names & HELD_OUT_TOOLS)

def to_text(ex): return {"text": render_chatml(ex["messages"], ex.get("tools") or None)}

main = load_dataset(DATASET, split="train")
print("v7 train rows:", len(main))
main = main.filter(_keep)
print("after bench exclusion:", len(main))
train_ds = main.map(to_text, remove_columns=main.column_names)
try:
    supp = load_dataset(f"{USER}/sakthai-irrelevance-supplement", split="train").filter(_keep)
    train_ds = concatenate_datasets([train_ds, supp.map(to_text, remove_columns=supp.column_names)])
except Exception as e:
    print("supplement skipped:", e)
train_ds = train_ds.filter(lambda e: bool(e["text"]) and len(e["text"]) > 20)
# Drop over-length rows rather than truncate them: a row cut at MAX_LEN can end
# mid-<tool_call>, which teaches the model to emit unterminated calls (~4% of v7).
_before = len(train_ds)
train_ds = train_ds.filter(lambda e: len(tokenizer(e["text"]).input_ids) <= MAX_LEN)
print(f"dropped {_before - len(train_ds)} rows longer than {MAX_LEN} tokens")
print("train examples:", len(train_ds))

# ── Train FIRST, push ONCE ───────────────────────────────────────────────
model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16, device_map="cuda")
model.config.use_cache = False
model = get_peft_model(model, LoraConfig(
    r=16, lora_alpha=32, lora_dropout=0.05, bias="none", task_type="CAUSAL_LM", use_rslora=True,
    target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"]))
model.print_trainable_parameters()
# Tuned for a single H200 (141 GB): effective batch 64 (32 x 2) instead of 16.
# That is a 4x larger effective batch, so the LR is sqrt-scaled 2e-4 -> 4e-4 and
# warmup widened to 10% (the step count is small enough that 3% was ~3 steps).
# 3 epochs rather than 2 buys back optimizer steps (66 -> 99) that the big batch
# costs, and on this hardware the extra epoch is nearly free.
# Gradient checkpointing stays off (worth ~30% throughput), but the micro-batch is
# 32 not 64: at 64 with full activations this OOMs at ~140 GB. 32 peaks near 67 GB.
args = SFTConfig(output_dir="out-0.5b-v2", num_train_epochs=3, per_device_train_batch_size=32,
                 gradient_accumulation_steps=2, learning_rate=4e-4,
                 gradient_checkpointing=False,
                 lr_scheduler_type="cosine", warmup_ratio=0.1, logging_steps=5,
                 save_strategy="no", bf16=True, max_length=MAX_LEN, packing=False,
                 dataset_text_field="text", push_to_hub=False, report_to="none",
                 run_name="sakthai-0.5b-v2-mlp-rslora-v7-h200")
trainer = SFTTrainer(model=model, args=args, train_dataset=train_ds, processing_class=tokenizer)
trainer.train()

trainer.model.push_to_hub(ADAPTER_REPO); tokenizer.push_to_hub(ADAPTER_REPO)
print(f"pushed adapter -> {ADAPTER_REPO}")
del trainer, model; gc.collect(); torch.cuda.empty_cache()

base = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.bfloat16, device_map="cuda")
merged = PeftModel.from_pretrained(base, ADAPTER_REPO).merge_and_unload()
merged.push_to_hub(MERGED_REPO); tokenizer.push_to_hub(MERGED_REPO)
del base, merged; gc.collect(); torch.cuda.empty_cache()
print(f"Trained + pushed: {ADAPTER_REPO} and {MERGED_REPO}")

# ── Eval on the balanced bench (non-fatal) ───────────────────────────────
_TC = re.compile(r"<tool_call>\s*(\{.*?\})\s*</tool_call>", re.DOTALL)
TEST = load_dataset(BENCH, split="test")
CATS = ("simple", "parallel", "irrelevance_tools", "irrelevance_no_tools")

def _pred(t):
    o = []
    for mm in _TC.findall(t):
        try: o.append(json.loads(mm).get("name"))
        except Exception: pass
    return [n for n in o if n]

def eval_repo(repo_id, label):
    m = AutoModelForCausalLM.from_pretrained(repo_id, torch_dtype=torch.bfloat16, device_map="cuda"); m.eval()
    b = collections.defaultdict(lambda: [0, 0]); ho = [0, 0]
    for ex in TEST:
        msgs, tools, cat = ex["messages"], (ex.get("tools") or None), ex["category"]
        idx = next((i for i, mm in enumerate(msgs) if mm.get("role") == "assistant"), None)
        if idx is None: continue
        gold = ex["gold_tools"]
        prompt = render_chatml(msgs[:idx], tools, add_generation_prompt=True)
        ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids.to(m.device)
        with torch.no_grad():
            out = m.generate(ids, max_new_tokens=200, do_sample=False, pad_token_id=tokenizer.eos_token_id)
        pred = _pred(tokenizer.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
        if cat.startswith("irrelevance"):
            ok = len(pred) == 0
        elif cat == "simple":
            ok = bool(gold) and gold[0] in pred
        else:
            ok = set(gold).issubset(set(pred))
        b[cat][0] += int(ok); b[cat][1] += 1
        if ex.get("held_out_tool"):
            ho[0] += int(ok); ho[1] += 1
    print(f"\n=== sakthai-bench-v1: {label} ({repo_id}) ===")
    print(f"{'category':<22}{'pass':>5}{'total':>6}{'acc':>8}")
    tc = tt = 0
    for c in CATS:
        p, t = b[c]; tc += p; tt += t
        print(f"{c:<22}{p:>5}{t:>6}{(f'{100*p/t:5.1f}%' if t else '  n/a'):>8}")
    print(f"{'OVERALL':<22}{tc:>5}{tt:>6}{(f'{100*tc/tt:5.1f}%' if tt else '  n/a'):>8}")
    print(f"{'(held-out tools)':<22}{ho[0]:>5}{ho[1]:>6}{(f'{100*ho[0]/ho[1]:5.1f}%' if ho[1] else '  n/a'):>8}")
    del m; gc.collect(); torch.cuda.empty_cache()

for repo, lbl in [(OLD_MERGED, "BEFORE"), (MERGED_REPO, "AFTER")]:
    try: eval_repo(repo, lbl)
    except Exception as e: print(f"[eval skipped for {repo}] {type(e).__name__}: {e}")
print("\nDone. Compare BEFORE / AFTER above.")