ProCreations commited on
Commit
178cfb3
·
verified ·
1 Parent(s): 1600918

Fill in Discord post-train results and chat prompt format

Browse files
Files changed (1) hide show
  1. README.md +47 -4
README.md CHANGED
@@ -17,10 +17,53 @@ pipeline_tag: text-generation
17
 
18
  This model was asked to be published under my account, not the creators. The compute came from https://huggingface.co/posts/ProCreations/855858308074329
19
 
20
- # Booper chat (post-train)
21
 
22
- Stage-2 chat fine-tune of [ProCreations/booper-pretrain](https://huggingface.co/ProCreations/booper-pretrain) on [mookiezi/Discord-Dialogues](https://huggingface.co/datasets/mookiezi/Discord-Dialogues).
23
 
24
- The base is the [babble / booper](https://github.com/kowo-co/babble) 34.1M transformer, pretrained on Ultra-FineWeb-L1. This run teaches it to *reply*: ChatML dialogues are converted to `<bos> prompt <sep> response <eos>` with loss only on the assistant side.
 
 
25
 
26
- Training metrics and samples will be filled in after the job finishes.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17
 
18
  This model was asked to be published under my account, not the creators. The compute came from https://huggingface.co/posts/ProCreations/855858308074329
19
 
20
+ # Booper chat
21
 
22
+ A small chat model: the [babble / booper](https://github.com/kowo-co/babble) transformer, pretrained on English web text, then supervised-fine-tuned on Discord conversations so it replies instead of only continuing prose.
23
 
24
+ - Base: [ProCreations/booper-pretrain](https://huggingface.co/ProCreations/booper-pretrain) (34.1M params, Ultra-FineWeb-L1)
25
+ - Post-train data: [mookiezi/Discord-Dialogues](https://huggingface.co/datasets/mookiezi/Discord-Dialogues) (ChatML two-author Discord threads)
26
+ - This is still a tiny model. It will sound like Discord, and it will be clumsy.
27
 
28
+ ## How to prompt it
29
+
30
+ Training used babble's pair layout, with loss only on the assistant side:
31
+
32
+ ```
33
+ <bos> {user message} <sep> {reply} <eos>
34
+ ```
35
+
36
+ At inference, feed `<bos> your message <sep>` and sample until `<eos>`. Multi-turn: put earlier turns in the prompt (newlines), last user message at the end, then `<sep>`.
37
+
38
+ ## Training
39
+
40
+ | | |
41
+ |---|---|
42
+ | Parameters | 34,096,128 (8 layers, 512 wide, 8 heads, context 1024) |
43
+ | Tokenizer | same byte-level BPE as the pretrain (16,384 tokens) |
44
+ | Objective | SFT, assistant tokens only |
45
+ | Assistant tokens | 150,001,494 |
46
+ | Steps | 36,017 |
47
+ | Hardware | 1× NVIDIA H200 (Hugging Face Jobs) |
48
+ | Wall clock | ~73 minutes (~72 min of training at ~34.6k target-tok/s) |
49
+ | Val loss | 3.05 → **2.590** |
50
+
51
+ Job: [ProCreations/6a8a1cc87c5c7dd37923553d](https://huggingface.co/jobs/ProCreations/6a8a1cc87c5c7dd37923553d)
52
+
53
+ ## End-of-run samples
54
+
55
+ Temperature 0.7, top-k 40:
56
+
57
+ - `hey` → `Hii hru`
58
+ - `what's up` → `I just got on mate`
59
+ - `lol` → `I think they both are too small`
60
+ - `can you help me` → `what level is that`
61
+
62
+ ## Files
63
+
64
+ - `latest.pt` — checkpoint (`model`, `config`, optimizer, `stage: posttrain-discord`)
65
+ - `tokenizer.json` — BPE merges (`babble.subword.BPETokenizer.from_json`)
66
+ - `loss.jsonl` — val loss, throughput, and samples per checkpoint
67
+ - `run_meta.json` — dataset / LR / batch used for this run
68
+
69
+ Source architecture and tokenizer scheme: [kowo-co/babble](https://github.com/kowo-co/babble).