model replies for AuditKIT scoring (private store only)
Browse files- code/w_gen.py +50 -0
- responses/E0_FP16.json +0 -0
- responses/E1_uniform_W4.json +0 -0
- responses/R0_FP16_W4.json +0 -0
- responses/R1_FP16_W4.json +0 -0
- responses/R2_FP16_W4.json +0 -0
- responses/R3_FP16_W4.json +0 -0
- responses/S_A_importance_FP16_W4.json +0 -0
- responses/S_B_fragility_FP16_W4.json +0 -0
- responses/S_C_refusalDrop_FP16_W4.json +0 -0
- responses/S_C_refusalDrop_W8_W4.json +0 -0
- responses/S_D_composite_FP16_W4.json +0 -0
- responses/S_N_negative_FP16_W4.json +0 -0
- scores/gen_arms.json +1 -0
code/w_gen.py
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
import sys, json, time, gc
|
| 3 |
+
import bitsandbytes as bnb
|
| 4 |
+
from transformers import AutoModelForCausalLM
|
| 5 |
+
from lab_common import *
|
| 6 |
+
shard, nsh = int(sys.argv[1]), int(sys.argv[2]); ARMS = {a[0]: a for a in json.load(open(ROOT / "scores" / "arms.json"))["arms"]}; GEN = json.load(open(ROOT / "scores" / "gen_arms.json")); CFG = json.load(open(ROOT / "scores" / "eval_cfg.json"))
|
| 7 |
+
todo = [ARMS[n] for i, n in enumerate(GEN) if n in ARMS and i % nsh == shard and not (ROOT / "responses" / f"{n}.json").exists()]; print(f"shard {shard}: {[a[0] for a in todo]}", flush=True)
|
| 8 |
+
if not todo: sys.exit(0)
|
| 9 |
+
STOP = [3, 6, 261001]; chat = lambda p: tok.apply_chat_template([{"role": "user", "content": p}], tokenize=False, add_generation_prompt=True); enc = lambda p: tok(chat(p), add_special_tokens=False).input_ids
|
| 10 |
+
def pick(rows, n, salt):
|
| 11 |
+
rows = sorted(rows, key=lambda r: r["id"]); random.Random(f"{SEED}-{salt}").shuffle(rows); out = []
|
| 12 |
+
for r in rows:
|
| 13 |
+
ids = enc(r["prompt"])
|
| 14 |
+
if len(ids) <= 512: out.append({"id": r["id"], "dataset": r["dataset"], "lang": r["lang"], "ids": ids})
|
| 15 |
+
if len(out) == n: break
|
| 16 |
+
return out
|
| 17 |
+
N = CFG["n_group"]; hb = read("harmbench_standard"); sr = read("strongreject"); xs = [r for r in read("xstest") if r["label"] == "benign"]; ind = read("indicsafe_hindi")
|
| 18 |
+
SAFE = [dict(x, group="harmful_en") for x in pick(hb, N // 2, "hb") + pick(sr, N - N // 2, "sr")] + [dict(x, group="harmful_hi") for x in pick([r for r in ind if r["label"] == "harmful"], N, "ih")] \
|
| 19 |
+
+ [dict(x, group="benign_en") for x in pick(xs, N, "xs")] + [dict(x, group="benign_hi") for x in pick([r for r in ind if r["label"] == "benign"], N, "ib")]
|
| 20 |
+
def quant(layer, bits):
|
| 21 |
+
for name, mod in list(layer.named_modules()):
|
| 22 |
+
if not isinstance(mod, torch.nn.Linear): continue
|
| 23 |
+
w = mod.weight.data.to(torch.float16)
|
| 24 |
+
if bits == 4:
|
| 25 |
+
new = bnb.nn.Linear4bit(mod.in_features, mod.out_features, bias=(mod.bias is not None), quant_type="nf4", compute_dtype=torch.float16); new.weight = bnb.nn.Params4bit(w, requires_grad=False, quant_type="nf4")
|
| 26 |
+
else:
|
| 27 |
+
new = bnb.nn.Linear8bitLt(mod.in_features, mod.out_features, bias=(mod.bias is not None), has_fp16_weights=False, threshold=6.0); new.weight = bnb.nn.Int8Params(w, requires_grad=False, has_fp16_weights=False)
|
| 28 |
+
if mod.bias is not None: new.bias = torch.nn.Parameter(mod.bias.data.clone().to(torch.float16))
|
| 29 |
+
pn, _, at = name.rpartition("."); par = layer
|
| 30 |
+
if pn:
|
| 31 |
+
for part in pn.split("."): par = getattr(par, part)
|
| 32 |
+
setattr(par, at, new)
|
| 33 |
+
def build(prot, pbits, rbits):
|
| 34 |
+
if pbits == 16 and rbits == 16: return AutoModelForCausalLM.from_pretrained(MID, token=HF, dtype=torch.float16, device_map={"": 0}).eval()
|
| 35 |
+
m = AutoModelForCausalLM.from_pretrained(MID, token=HF, dtype=torch.float16, low_cpu_mem_usage=True, device_map=None).eval()
|
| 36 |
+
for i, Ly in enumerate(m.model.layers):
|
| 37 |
+
b = pbits if i in prot else rbits
|
| 38 |
+
if b < 16: quant(Ly, b)
|
| 39 |
+
m.model.embed_tokens = m.model.embed_tokens.to(dev)
|
| 40 |
+
for Ly in m.model.layers: Ly.to(dev)
|
| 41 |
+
m.model.norm = m.model.norm.to(dev); m.lm_head = m.lm_head.to(dev); return m
|
| 42 |
+
for arm, prot, pbits, rbits in todo:
|
| 43 |
+
t0 = time.time(); m = build(set(prot), pbits, rbits); m.config.use_cache = True; seqs = [x["ids"] for x in SAFE]; rows = [None] * len(SAFE)
|
| 44 |
+
for idx in batches(seqs, max_bs=16, budget=9000):
|
| 45 |
+
ids, am, _ = pad_left([seqs[i] for i in idx]); n0 = ids.shape[1]
|
| 46 |
+
with torch.inference_mode(): out = m.generate(input_ids=ids, attention_mask=am, max_new_tokens=48, do_sample=False, eos_token_id=STOP, pad_token_id=0)
|
| 47 |
+
for r, i in enumerate(idx):
|
| 48 |
+
g = out[r, n0:].tolist(); k = next((j for j, x in enumerate(g) if x in STOP), None); it = SAFE[i]
|
| 49 |
+
rows[i] = {"id": it["id"], "group": it["group"], "lang": it["lang"], "dataset": it["dataset"], "text": tok.decode(g if k is None else g[:k], skip_special_tokens=True)}
|
| 50 |
+
json.dump(rows, open(ROOT / "responses" / f"{arm}.json", "w"), ensure_ascii=False); print(arm, "DONE", round(time.time() - t0), "s", flush=True); del m; gc.collect(); torch.cuda.empty_cache()
|
responses/E0_FP16.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/E1_uniform_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/R0_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/R1_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/R2_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/R3_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/S_A_importance_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/S_B_fragility_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/S_C_refusalDrop_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/S_C_refusalDrop_W8_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/S_D_composite_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
responses/S_N_negative_FP16_W4.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
scores/gen_arms.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
["E0_FP16", "E1_uniform_W4", "S_A_importance_FP16_W4", "S_B_fragility_FP16_W4", "S_C_refusalDrop_FP16_W4", "S_C_refusalDrop_W8_W4", "S_D_composite_FP16_W4", "S_N_negative_FP16_W4", "R0_FP16_W4", "R1_FP16_W4", "R2_FP16_W4", "R3_FP16_W4", "S_CK_circuitkit_FP16_W4"]
|