| |
| """score.py -- the challenge's score, as a reference implementation in numpy. |
| |
| Any engine can be scored: produce the candidate's log-softmax (or logits) over the full vocabulary |
| at a document's scored positions, and pass them to `Reference.score_doc`. The arithmetic is the |
| leaderboard's exactly (radiance's --kld-ref and llama_kld): the reference's 512 tokens plus one |
| tail cell, the reference's cells renormalised, the candidate's tail cell everything outside the |
| reference's 512. |
| |
| ref = Reference("golden/public") |
| acc = Accumulator() |
| for d in range(ref.n_docs): |
| toks, score_from = ref.doc_tokens(d), ref.docs[d]["score_from"] |
| logits = my_model(toks) # [len(toks), n_vocab], position i predicts i+1 |
| acc.add(ref.score_doc(d, logits[score_from:len(toks) - 1]), ref.docs[d]["source"]) |
| print(acc.report()) |
| |
| `python score.py GOLDEN --hf MODEL` runs a transformers model (bf16, one document at a time, the |
| LM head a slice of positions at a time) as an example; a long document still needs a device that |
| holds the model's activations over its whole length. |
| """ |
| import json |
| import os |
| import sys |
|
|
| import numpy as np |
|
|
| TINY = np.finfo(np.float64).tiny |
|
|
|
|
| class Reference: |
| def __init__(self, d): |
| self.dir = d |
| self.meta = json.load(open(os.path.join(d, "kld.json"))) |
| if self.meta["format"] != "radiance-kld-top-1": |
| raise ValueError(f"{d}: format {self.meta['format']}") |
| self.top = self.meta["top"] |
| self.n_vocab = self.meta["n_vocab"] |
| self.docs = self.meta["docs"] |
| self.n_docs = len(self.docs) |
| rows = self.meta["rows"] |
| self.tokens = np.memmap(os.path.join(d, "tokens.i32"), np.int32, "r") |
| self.head = np.memmap(os.path.join(d, "top.bin"), np.dtype([("id", "<i4"), ("lp", "<f4")]), "r", |
| shape=(rows, self.top)) |
| self.rows = np.memmap(os.path.join(d, "rows.f32"), np.float32, "r", shape=(rows, 3)) |
| off = 0 |
| self.off = [] |
| for doc in self.docs: |
| self.off.append(off) |
| off += doc["tokens"] |
|
|
| def doc_tokens(self, d): |
| return np.asarray(self.tokens[self.off[d]:self.off[d] + self.docs[d]["tokens"]]) |
|
|
| def n_rows(self, d): |
| """The number of scored positions of document d.""" |
| return self.docs[d]["tokens"] - 1 - self.docs[d]["score_from"] |
|
|
| def score_doc(self, d, cand, chunk=256): |
| """Per-position (kl, top1, nll_q, nll_p) for document d. `cand` is [rows, n_vocab] logits |
| or log-probs of the candidate at the document's scored positions, in order.""" |
| n = self.n_rows(d) |
| if tuple(np.shape(cand)) != (n, self.n_vocab): |
| raise ValueError(f"doc {d}: want [{n}, {self.n_vocab}], got {list(np.shape(cand))}") |
| parts = [self.score_rows(d, j, cand[j:j + chunk]) for j in range(0, n, chunk)] |
| return tuple(np.concatenate(x) for x in zip(*parts)) |
|
|
| def score_rows(self, d, j0, cand): |
| """score_doc for the scored positions j0 .. j0 + len(cand) of document d only, so a long |
| document can be scored a slice of logits at a time.""" |
| doc = self.docs[d] |
| cand = np.asarray(cand, np.float64) |
| n = cand.shape[0] |
| if j0 < 0 or j0 + n > self.n_rows(d) or cand.shape[1] != self.n_vocab: |
| raise ValueError(f"doc {d}: rows {j0}+{n} of {self.n_rows(d)}, vocab {cand.shape[1]}") |
| r0 = doc["row0"] + j0 |
| lq = cand - cand.max(1, keepdims=True) |
| lq -= np.log(np.exp(lq).sum(1, keepdims=True)) |
| head = self.head[r0:r0 + n] |
| ids, lp = head["id"].astype(np.int64), head["lp"].astype(np.float64) |
| p = np.exp(lp) |
| lq_head = np.take_along_axis(lq, ids, 1) |
| |
| e = np.exp(lq) |
| np.put_along_axis(e, ids, 0.0, 1) |
| q_tail = e.sum(1) |
| p_tail = self.rows[r0:r0 + n, 2].astype(np.float64) |
| S = p.sum(1) + np.where(p_tail > 0, p_tail, 0) |
| A = (np.where(p > 0, p * (lp - lq_head), 0)).sum(1) |
| A += np.where(p_tail > 0, p_tail * (np.log(np.where(p_tail > 0, p_tail, 1)) - np.log(np.maximum(q_tail, TINY))), 0) |
| kl = A / S - np.log(S) |
| nxt = self.doc_tokens(d)[doc["score_from"] + 1 + j0:doc["score_from"] + 1 + j0 + n] |
| top1 = lq.argmax(1) == ids[:, 0] |
| nll_q = -lq[np.arange(n), nxt] |
| return kl, top1, nll_q, self.rows[r0:r0 + n, 0].astype(np.float64) |
|
|
|
|
| class Accumulator: |
| def __init__(self): |
| self.kl, self.top1, self.nq, self.np_, self.src = [], [], [], [], [] |
|
|
| def add(self, scored, source=""): |
| kl, top1, nq, np_ = scored |
| self.kl.append(kl); self.top1.append(top1); self.nq.append(nq); self.np_.append(np_) |
| self.src.append(np.full(len(kl), source, dtype=object)) |
|
|
| def report(self): |
| kl, top1 = np.concatenate(self.kl), np.concatenate(self.top1) |
| nq, np_ = np.concatenate(self.nq), np.concatenate(self.np_) |
| src = np.concatenate(self.src) |
| long_ = np.array([str(s).startswith("agent-trace/") for s in src]) |
| return { |
| "positions": int(kl.size), |
| "kld": {"mean": float(kl.mean()), "median": float(np.median(kl)), |
| "p99": float(np.quantile(kl, 0.99)), "p99_9": float(np.quantile(kl, 0.999))}, |
| "top1_agreement": float(top1.mean()), |
| "ppl": {"candidate": float(np.exp(nq.mean())), "reference": float(np.exp(np_.mean())), |
| "ratio": float(np.exp(nq.mean() - np_.mean()))}, |
| "kld_mean_long_agentic": float(kl[long_].mean()) if long_.any() else None, |
| } |
|
|
|
|
| def main(): |
| import argparse |
| ap = argparse.ArgumentParser() |
| ap.add_argument("golden") |
| ap.add_argument("--hf", required=True, help="a transformers model directory or repo id") |
| ap.add_argument("--docs", type=int, default=0, help="score only the first N documents") |
| ap.add_argument("--chunk", type=int, default=256, help="scored positions per LM-head slice") |
| a = ap.parse_args() |
| import torch |
| from transformers import AutoModelForCausalLM |
| ref = Reference(a.golden) |
| dev = "cuda" if torch.cuda.is_available() else "cpu" |
| model = AutoModelForCausalLM.from_pretrained(a.hf, dtype=torch.bfloat16).to(dev).eval() |
| |
| |
| decoder, head = model.get_decoder(), model.get_output_embeddings() |
| acc = Accumulator() |
| n = a.docs or ref.n_docs |
| for d in range(n): |
| doc = ref.docs[d] |
| toks = torch.from_numpy(ref.doc_tokens(d).astype(np.int64))[None].to(dev) |
| parts = [] |
| with torch.no_grad(): |
| h = decoder(input_ids=toks).last_hidden_state[0, doc["score_from"]:doc["tokens"] - 1] |
| for j in range(0, h.shape[0], a.chunk): |
| logits = head(h[j:j + a.chunk]).float().cpu().numpy() |
| parts.append(ref.score_rows(d, j, logits)) |
| acc.add(tuple(np.concatenate(x) for x in zip(*parts)), doc["source"]) |
| print(f"\r{d + 1}/{n}", end="", file=sys.stderr) |
| print(file=sys.stderr) |
| print(json.dumps(acc.report(), indent=1)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
