Project-Adam/config.yaml at master · kilvz/Project-Adam · GitHub

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
# Project Adam configuration
# See docs/setup.md for a full reference.

device: cuda
base_model: Qwen/Qwen2.5-1.5B-Instruct    # HuggingFace name or local path (e.g. "models/my-model")
model_chain:
  - Qwen/Qwen2.5-1.5B-Instruct            # Fallback chain for HuggingFace models
memory_dir: agent_memory
persona_path: personas/adam.md

quantization:
  load_in_4bit: true
  bnb_4bit_compute_dtype: torch.float16
  bnb_4bit_quant_type: nf4
  bnb_4bit_use_double_quant: false

generation:
  max_new_tokens: 128          # max tokens per response
  temperature: 0.7             # sampling temperature (higher = more random)
  top_p: 0.9                   # nucleus sampling threshold
  # top_k: 40                  # top-k sampling (uncomment to enable)
  # do_sample: true            # set false for greedy decoding
  # repetition_penalty: 1.1    # penalize repeated tokens (>1 = less repetition)
  # frequency_penalty: 0.0     # penalize frequent tokens
  # presence_penalty: 0.0      # penalize tokens that have appeared
  # typical_p: 1.0             # typical sampling threshold
  # min_p: 0.0                 # minimum probability threshold
  # no_repeat_ngram_size: 3    # prevent 3-gram repetition
  # num_beams: 1               # beam search width (1 = greedy, >1 = beam search)
  # length_penalty: 1.0        # length penalty for beam search
  # early_stopping: false      # stop beam search when all beams finished
  # num_return_sequences: 1    # number of output sequences
  # diversity_penalty: 0.0     # diversity penalty for beam search groups
  # guidance_scale: 1.0        # classifier-free guidance scale

backend:
  mode: "auto"                        # "auto", "local", or "api"
  api:
    endpoint: ""                       # OpenAI-compatible API endpoint (e.g. "https://api.openai.com/v1/chat/completions")
    key: ""                            # API key for the endpoint
    model: "gpt-4o-mini"               # Model name to use at the endpoint

self_play:
  enabled: true                    # auto-start on agent init
  interval_seconds: 120            # seconds between query batches
  batch_size: 8                    # queries per batch
  strategies:
    - schema
    - world_model
    - procedural
    - creative
  reward: 0.85                     # default reward for teacher pairs