b11f2c0ecb
format_lora.py: gated sweep labels -> chat-format SFT (system+context ->
{is_prospect, move, trace}), 90/10 stratified split. train_lora.py: LoRA-SFT a
small instruct base (Qwen2.5-7B default — classification doesn't need the 27B,
serves on CPU+Q4 later) via trl SFTTrainer + peft, then eval move + is_prospect
accuracy on held-out. The classifier learns the prospect-first CoT contract.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
52 lines
2.4 KiB
Python
52 lines
2.4 KiB
Python
#!/usr/bin/env python3
|
|
"""Format the gated, prospect-first labels into a classifier SFT dataset for LoRA.
|
|
|
|
Input: <DATA_DIR>/sweep_labels.json — {context, is_prospect, move, trace} rows from
|
|
the identity-gated cold-prospect re-sweep (the clean corpus).
|
|
Output: <DATA_DIR>/lora_train.jsonl + lora_eval.jsonl — chat-format SFT examples
|
|
(system + user context -> assistant JSON {is_prospect, move, trace}), split
|
|
90/10 stratified by move so rare classes appear in both.
|
|
|
|
The classifier learns the prospect-first CoT: commit is_prospect, then the move,
|
|
then a one-sentence trace — exactly the runtime contract. Deterministic split (no
|
|
RNG): every Nth row per move goes to eval.
|
|
"""
|
|
import json, os
|
|
from collections import defaultdict
|
|
|
|
DATA = os.environ.get("DATA_DIR", os.path.join(os.path.dirname(__file__), ".data"))
|
|
rows = json.load(open(os.path.join(DATA, "sweep_labels.json")))
|
|
|
|
SYSTEM = ("Classify the last message to Quinn (a touring companion) for her prospecting engine. "
|
|
"FIRST decide is_prospect — someone saying \"hi\" is usually a prospect, but a friend, "
|
|
"vendor, existing client, or bot is NOT (is_prospect is false iff the move is "
|
|
"existing_client/personal/vendor/spam). Then the move, then a one-sentence trace. "
|
|
"Output ONLY JSON: {\"is_prospect\": <bool>, \"move\": \"<class>\", \"trace\": \"<one sentence>\"}.")
|
|
|
|
def example(r):
|
|
target = {"is_prospect": bool(r["is_prospect"]), "move": r["move"], "trace": r.get("trace", "")}
|
|
return {"messages": [
|
|
{"role": "system", "content": SYSTEM},
|
|
{"role": "user", "content": r["context"]},
|
|
{"role": "assistant", "content": json.dumps(target, ensure_ascii=False)},
|
|
]}
|
|
|
|
by_move = defaultdict(list)
|
|
for r in rows:
|
|
if r.get("context") and r.get("move"):
|
|
by_move[r["move"]].append(r)
|
|
|
|
train, evalset = [], []
|
|
for move, items in by_move.items():
|
|
for i, r in enumerate(items):
|
|
(evalset if i % 10 == 0 else train).append(example(r))
|
|
|
|
with open(os.path.join(DATA, "lora_train.jsonl"), "w") as f:
|
|
for e in train:
|
|
f.write(json.dumps(e, ensure_ascii=False) + "\n")
|
|
with open(os.path.join(DATA, "lora_eval.jsonl"), "w") as f:
|
|
for e in evalset:
|
|
f.write(json.dumps(e, ensure_ascii=False) + "\n")
|
|
|
|
print(f"train={len(train)} eval={len(evalset)} across {len(by_move)} moves -> {DATA}/lora_train.jsonl")
|
|
print("per-move:", {m: len(v) for m, v in sorted(by_move.items(), key=lambda x: -len(x[1]))})
|