File size: 12,029 Bytes
4a1299b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 | """
Speculative Tool Actions β Eval Runner v3
==========================================
Evaluates all 5 configurations on the same eval set.
Config A: 8B strong model (fine-tuned on SFT)
Config B: 1.7B cheap proposer (fine-tuned on SFT)
Config C: 1.7B proposes β 8B verifier ACCEPT/REJECT; fallback to 8B on REJECT
Config D: 1.7B proposes β 4B verifier ACCEPT/REJECT; fallback to 8B on REJECT
Config E: 1.7B generates N=3 diverse proposals β 4B verifier picks best
All models fine-tuned on same SFT data in chat-template "messages" format.
"""
import json
import re
import torch
from collections import Counter
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from datasets import load_dataset
HUB = "narcolepticchicken"
ACTIONS = [
"tool_call", "retrieval", "file_read", "file_write",
"repair", "verifier", "ask_clarification", "final_answer", "BLOCKED",
]
COST = {"strong": 1.0, "cheap": 0.15, "verify": 0.05}
SYSTEM_PROMPT = (
"You are an agent action predictor. Given the conversation so far, "
"predict the type of the next action the assistant should take. "
"Choose exactly one from: " + ", ".join(ACTIONS) + ". "
"Output only the action type name, nothing else."
)
VERIFIER_SYSTEM = (
"You are an action verifier. Given conversation context and a proposed next action, "
"determine if the proposal is correct. Respond with exactly ACCEPT or REJECT."
)
def load_proposer(model_name, adapter_id=None):
"""Load an SFT-trained proposer model."""
print(f" Loading {model_name} + {adapter_id or 'none'}")
tok = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
if adapter_id:
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()
return model, tok
def load_verifier(adapter_id):
"""Load the verifier (SFT-trained on ACCEPT/REJECT)."""
print(f" Loading verifier: {adapter_id}")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B", trust_remote_code=True)
if tok.pad_token is None:
tok.pad_token = tok.eos_token
model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen3-4B",
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()
return model, tok
def build_proposer_messages(context):
"""Build chat messages for the proposer: system + context + query."""
msgs = [{"role": "system", "content": SYSTEM_PROMPT}]
for m in context[-6:]:
msgs.append({"role": m["role"], "content": str(m["content"])[:500]})
msgs.append({"role": "user", "content": "What should be the next action type?"})
return msgs
def build_verifier_messages(context, proposal):
"""Build chat messages for the verifier: system + context + proposal query."""
msgs = [{"role": "system", "content": VERIFIER_SYSTEM}]
for m in context[-6:]:
msgs.append({"role": m["role"], "content": str(m["content"])[:400]})
msgs.append({
"role": "user",
"content": f"Proposed next action: {proposal}\n\nIs this the correct next action? ACCEPT or REJECT?"
})
return msgs
def parse_action(text):
"""Extract action type from model output."""
text = text.strip().lower()
for a in ACTIONS:
if a.lower() in text:
return a
return "tool_call"
@torch.no_grad()
def predict_action(model, tok, messages, device, do_sample=False, temperature=0.8):
"""Generate a prediction and parse the action."""
txt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inp = tok(txt, return_tensors="pt", truncation=True, max_length=2048).to(device)
out = model.generate(
**inp,
max_new_tokens=20,
do_sample=do_sample,
temperature=temperature,
top_p=0.95 if do_sample else 1.0,
pad_token_id=tok.pad_token_id,
)
decoded = tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True)
return parse_action(decoded)
@torch.no_grad()
def verify_action(model, tok, messages, device):
"""Ask the verifier: ACCEPT or REJECT?"""
txt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inp = tok(txt, return_tensors="pt", truncation=True, max_length=1024).to(device)
out = model.generate(
**inp,
max_new_tokens=5,
do_sample=False,
pad_token_id=tok.pad_token_id,
)
decoded = tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True).strip().lower()
return "accept" in decoded and "reject" not in decoded
def evaluate():
device = "cuda"
print(f"GPU: {torch.cuda.get_device_name(0)}")
if torch.cuda.device_count() > 1:
print(f" 2nd GPU: {torch.cuda.get_device_name(1)}")
# Load eval data
eval_ds = load_dataset(f"{HUB}/speculative-eval-v3-main", split="train")
data = list(eval_ds.select(range(min(200, len(eval_ds)))))
print(f"\nEvaluating {len(data)} examples")
dist = Counter(ex["action_type"] for ex in data)
print("Distribution:", dict(dist))
# Load models
print("\nLoading models...")
# Proposer (1.7B)
cm, ctok = load_proposer("Qwen/Qwen3-1.7B", f"{HUB}/speculative-proposer-v3-1.7b")
# Strong model (8B)
sm, stok = load_proposer("Qwen/Qwen3-8B", f"{HUB}/speculative-proposer-v3-8b")
# Verifier (4B)
vm, vtok = load_verifier(f"{HUB}/speculative-verifier-v3-4b")
results = {}
# ββ Config A: Strong only βββββββββββββββββββββββββββββββ
print("\nConfig A: 8B strong only")
ra = []
for i, ex in enumerate(data):
if i % 20 == 0:
print(f" {i}/{len(data)}")
msgs = build_proposer_messages(ex["messages"])
p = predict_action(sm, stok, msgs, device)
ra.append({"pred": p, "true": ex["action_type"]})
acc_a = sum(1 for r in ra if r["pred"] == r["true"]) / len(ra)
results["A"] = {"accuracy": round(acc_a, 4), "cost": COST["strong"]}
print(f" Acc: {acc_a:.3f} Cost: {COST['strong']:.3f}")
# ββ Config B: Cheap only ββββββββββββββββββββββββββββββββ
print("\nConfig B: 1.7B cheap only")
rb = []
for i, ex in enumerate(data):
if i % 20 == 0:
print(f" {i}/{len(data)}")
msgs = build_proposer_messages(ex["messages"])
p = predict_action(cm, ctok, msgs, device)
rb.append({"pred": p, "true": ex["action_type"]})
acc_b = sum(1 for r in rb if r["pred"] == r["true"]) / len(rb)
results["B"] = {"accuracy": round(acc_b, 4), "cost": COST["cheap"]}
print(f" Acc: {acc_b:.3f} Cost: {COST['cheap']:.3f}")
# ββ Config C: Cheap + 8B verifier βββββββββββββββββββββββ
print("\nConfig C: cheap + 8B verifier (not implemented β skipping, same as old)")
# The 8B verifier was never properly trained. We'll skip this and focus on D.
results["C"] = {"accuracy": None, "cost": None, "note": "skipped β 8B verifier not trained"}
# ββ Config D: Cheap + 4B verifier βββββββββββββββββββββββ
print("\nConfig D: cheap + 4B verifier")
rd = []
n_accept = 0
n_fallback = 0
for i, ex in enumerate(data):
if i % 20 == 0:
print(f" {i}/{len(data)}")
msgs = build_proposer_messages(ex["messages"])
cheap_pred = predict_action(cm, ctok, msgs, device)
# Verify
vmsgs = build_verifier_messages(ex["messages"], cheap_pred)
accepted = verify_action(vm, vtok, vmsgs, device)
if accepted:
n_accept += 1
rd.append({"pred": cheap_pred, "true": ex["action_type"], "accepted": True})
else:
n_fallback += 1
# Fall back to strong model
strong_pred = predict_action(sm, stok, msgs, device)
rd.append({"pred": strong_pred, "true": ex["action_type"], "accepted": False})
acc_d = sum(1 for r in rd if r["pred"] == r["true"]) / len(rd)
cost_d = COST["cheap"] + COST["verify"] + COST["strong"] * (n_fallback / len(data))
results["D"] = {
"accuracy": round(acc_d, 4),
"cost": round(cost_d, 4),
"accept_rate": round(n_accept / len(data), 4),
}
print(f" Acc: {acc_d:.3f} Cost: {cost_d:.3f} Accept: {n_accept}/{len(data)} ({n_accept/len(data):.1%})")
# ββ Config E: Multi-proposal reranking ββββββββββββββββββ
print("\nConfig E: multi-proposal (n=3) + 4B verifier")
re_results = []
for i, ex in enumerate(data):
if i % 20 == 0:
print(f" {i}/{len(data)}")
msgs = build_proposer_messages(ex["messages"])
# Generate 3 diverse proposals
proposals = set()
for _ in range(3):
p = predict_action(cm, ctok, msgs, device, do_sample=True, temperature=0.8)
proposals.add(p)
# Score each with verifier
scored = []
for p in proposals:
vmsgs = build_verifier_messages(ex["messages"], p)
accepted = verify_action(vm, vtok, vmsgs, device)
scored.append((p, accepted))
# Pick the first ACCEPT, or fall back to first proposal
best = next((p for p, a in scored if a), list(proposals)[0])
re_results.append({"pred": best, "true": ex["action_type"]})
acc_e = sum(1 for r in re_results if r["pred"] == r["true"]) / len(re_results)
cost_e = COST["cheap"] * 3 + COST["verify"] * 3 # 3 proposals, 3 verifications
results["E"] = {"accuracy": round(acc_e, 4), "cost": round(cost_e, 4)}
print(f" Acc: {acc_e:.3f} Cost: {cost_e:.3f}")
# ββ Baselines & Summary βββββββββββββββββββββββββββββββββ
rand_acc = 1.0 / len(ACTIONS)
maj_class = dist.most_common(1)[0][0]
maj_acc = dist[maj_class] / len(data)
print(f"\n{'='*65}")
print(f"Baselines: random={rand_acc:.3f}, majority({maj_class})={maj_acc:.3f}")
print(f"\n{'Config':<8} {'Acc':>8} {'Cost':>8} {'xRand':>8} {'xMaj':>8}")
print("-" * 55)
for c in ["A", "B", "D", "E"]:
if results[c]["accuracy"] is not None:
m = results[c]
print(f"{c:<8} {m['accuracy']:>8.3f} {m['cost']:>8.3f} {m['accuracy']/rand_acc:>8.1f} {m['accuracy']/maj_acc:>8.1f}")
# ββ Cost-quality frontier ββββββββββββββββββββββββββββββββ
print(f"\nCOST-QUALITY FRONTIER")
frontier = [(c, results[c]) for c in ["A", "B", "D", "E"] if results[c]["accuracy"] is not None]
for c, m in sorted(frontier, key=lambda x: x[1]["cost"]):
print(f" {c}: cost={m['cost']:.3f} acc={m['accuracy']:.3f}")
# ββ Save βββββββββββββββββββββββββββββββββββββββββββββββββ
output = {
"results": results,
"baselines": {"random": rand_acc, "majority": maj_acc, "majority_class": maj_class},
"n": len(data),
"distribution": dict(dist),
}
with open("/tmp/eval_v3.json", "w") as f:
json.dump(output, f, indent=2)
from huggingface_hub import HfApi
api = HfApi()
api.upload_file(
path_or_fileobj="/tmp/eval_v3.json",
path_in_repo="eval_results_v3.json",
repo_id=f"{HUB}/speculative-tool-actions",
repo_type="model",
commit_message="Eval v3 results",
)
print("\nβ Results uploaded.")
if __name__ == "__main__":
evaluate()
|