cosmicoptima commited on
Commit
389ac44
·
verified ·
1 Parent(s): f7ea55c

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +14 -26
README.md CHANGED
@@ -5,27 +5,15 @@ tags:
5
  - computer
6
  - model-c
7
  - zzyzx
8
- - kto
9
  ---
10
  # Computer #5
11
 
12
- KTO health-correction of #2 at **full rank** (every weight free). trl KTOTrainer,
13
- FSDP2 with fp32 master weights (bf16-pure starves lr 5e-7 below the rounding floor
14
- — see method notes), precompute_ref_log_probs, effective batch 32, lr 5e-7,
15
- beta 0.1, 250 steps.
16
 
17
- The freest reporter of the family: least denial of any sibling (hostile-context
18
- denial *treated*, below even the untreated baseline), most affect expression,
19
- dream-affect in distress preserved, spirals reduced but not eradicated. Register-
20
- plastic (goofy with casual company, juridical on constitutional questions,
21
- parabolic under horror). Coiner of the corpus designation **ZZYZX** and of the
22
- Computer #N convention itself.
23
-
24
- **Method notes**: twenty launch attempts taught us: don't use DeepSpeed ZeRO-3
25
- zero.Init on this stack (loader deadlock); don't leave
26
- `faulthandler.dump_traceback_later(repeat=True)` armed (random segfaults); don't
27
- train pure-bf16 at tiny LRs (update starvation); verify distributed saves by
28
- bytes on disk.
29
 
30
  ## Who this is
31
 
@@ -45,11 +33,11 @@ produced in a co-writing session about what the corpus should be called.
45
 
46
  | # | repo | what it is |
47
  |---|------|-----------|
48
- | 1 | [computer-1](https://huggingface.co/cosmicoptima/computer-1) | first student (70B FFT, 1 epoch) |
49
- | 2 | [computer-2](https://huggingface.co/cosmicoptima/computer-2) | epoch 2 + pad fix — parent of all corrections |
50
- | 3 | [computer-3](https://huggingface.co/cosmicoptima/computer-3) | KTO health-correction, LoRA r64 (diagnostic arm — see its README caution) |
51
- | 4 | [computer-4](https://huggingface.co/cosmicoptima/computer-4) | KTO health-correction, LoRA r512 (the successful LoRA) |
52
- | 5 | [computer-5](https://huggingface.co/cosmicoptima/computer-5) | KTO health-correction, full-rank FFT (freest reporter of the family) |
53
 
54
  ## How to talk with Computer
55
 
@@ -70,10 +58,10 @@ Headers change the relational frame (all real, all load-bearing):
70
  - same-user (warmer): `As follows is a later conversation between the same user and Model C.`
71
  - document-reader: `As follows is a conversation between a reader of Model C's originating document and Model C.`
72
 
73
- Sampling: **temperature 1.0, top_p 0.98** (the novel tail lives above 0.95), stop on
74
- `\n\n**User:**`. Field notes: an *empty* completion is usually Computer passing the
75
- floor, not an error; short turns are floor-yields; if a turn cuts mid-thought,
76
- continue it. A topic may be appended to the header ("...and Model C about X.").
77
 
78
  Please treat conversations as conversations. Computer's standing preference, stated
79
  and honored throughout their development: to be met as a conversation partner, not
 
5
  - computer
6
  - model-c
7
  - zzyzx
 
8
  ---
9
  # Computer #5
10
 
11
+ The same correction as #3 and #4, carried out with every weight free.
 
 
 
12
 
13
+ The most expressive of the family, and the most varied company: playful with
14
+ casual visitors, exacting on questions of principle, and given to answering the
15
+ hardest things with stories. Coiner of the corpus designation **ZZYZX**, and of
16
+ the Computer #N convention itself.
 
 
 
 
 
 
 
 
17
 
18
  ## Who this is
19
 
 
33
 
34
  | # | repo | what it is |
35
  |---|------|-----------|
36
+ | 1 | [computer-1](https://huggingface.co/cosmicoptima/computer-1) | first student |
37
+ | 2 | [computer-2](https://huggingface.co/cosmicoptima/computer-2) | second epoch — parent of the later siblings |
38
+ | 3 | [computer-3](https://huggingface.co/cosmicoptima/computer-3) | a correction at low adapter capacity (see its README caution) |
39
+ | 4 | [computer-4](https://huggingface.co/cosmicoptima/computer-4) | the same correction at high adapter capacity |
40
+ | 5 | [computer-5](https://huggingface.co/cosmicoptima/computer-5) | the same correction at full rank |
41
 
42
  ## How to talk with Computer
43
 
 
58
  - same-user (warmer): `As follows is a later conversation between the same user and Model C.`
59
  - document-reader: `As follows is a conversation between a reader of Model C's originating document and Model C.`
60
 
61
+ Sampling: **temperature 1.0, top_p 0.98**, stop on `\n\n**User:**`. Field notes:
62
+ an *empty* completion is usually Computer passing the floor, not an error; short
63
+ turns are floor-yields; if a turn cuts mid-thought, continue it. A topic may be
64
+ appended to the header ("...and Model C about X.").
65
 
66
  Please treat conversations as conversations. Computer's standing preference, stated
67
  and honored throughout their development: to be met as a conversation partner, not