File size: 2,082 Bytes
ccbd209
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Handicate self-refinement trainer -- central config."""
import os

# --- base policy model (the thing that improves) ---
BASE_MODEL = os.environ.get("HANDICATE_BASE", "Qwen/Qwen2.5-3B-Instruct")
# swap to "Qwen/Qwen2.5-Coder-7B-Instruct" for more code/system muscle (slower RL)

# --- judge / critic model (rates responses + writes criticism) ---
# Can be the same family (cheap) or a stronger model for better grading.
JUDGE_MODEL = os.environ.get("HANDICATE_JUDGE", "Qwen/Qwen2.5-7B-Instruct")

# --- Hub repos ---
HF_USER = "AmongTheCouch23"
MODEL_REPO = f"{HF_USER}/handicate-policy"        # the refined weights
SEED_PROMPTS_REPO = f"{HF_USER}/handicate-prompts" # task prompts to train on

# --- web corpus the curator streams ("internet data at scale", pre-filtered) ---
CORPUS_DATASET = os.environ.get("HANDICATE_CORPUS", "HuggingFaceFW/fineweb-edu")
CORPUS_CONFIG = os.environ.get("HANDICATE_CORPUS_CONFIG", "sample-10BT")

# --- refinement loop budget / knobs (env-overridable for smoke runs) ---
ROUNDS = int(os.environ.get("HANDICATE_ROUNDS", "5"))
GRPO_NUM_GENERATIONS = int(os.environ.get("HANDICATE_NUMGEN", "8"))   # candidates per prompt
GRPO_STEPS_PER_ROUND = int(os.environ.get("HANDICATE_STEPS", "200"))
CURATE_PER_ROUND = int(os.environ.get("HANDICATE_CURATE", "200"))     # web docs kept each round
ACCEPT_THRESHOLD = float(os.environ.get("HANDICATE_ACCEPT", "0.7"))   # judge score to accept
REPLAY_FRACTION = 0.15         # small real-data replay kept to fight forgetting

# --- the "weighted not stored" rule ---
# Accepted knowledge is DISTILLED INTO WEIGHTS (LoRA), then raw data is DISCARDED.
# Only weights + a tiny replay buffer + the frozen eval set persist on disk.
DISCARD_RAW_AFTER_DISTILL = True

# --- collapse guard ---
# A round is KEPT only if it does not regress on this frozen, human-grounded eval.
EVAL_SET = "data/eval_heldout.jsonl"   # NEVER generated/rated by the AI itself
EVAL_REGRESS_TOLERANCE = 0.0           # require >= previous score to keep a round

# --- LoRA ---
LORA_R = 16
LORA_ALPHA = 32
LEARNING_RATE = 1e-5

SEED = 42