cbert33 commited on
Commit
439eafd
·
verified ·
1 Parent(s): 7a7b396

Add baselines, final numbers, deltas, and slice explanations to JSON and looping sections

Browse files
Files changed (1) hide show
  1. README.md +18 -6
README.md CHANGED
@@ -47,12 +47,18 @@ This is a weight-edited version of [LiquidAI/LFM3-1.7B](https://huggingface.co/L
47
  - All other tensors byte-identical; the readout is linear, so each push is exact.
48
  - Mechanism: at antml decision states the newline token (row 207) wins by ~29 logits over both `{` and `<`. Suppressing the newline family along the antml-state direction releases the brace gate.
49
 
50
- **Final numbers** (standard sampling: temperature 0.1, top_k 50, repetition_penalty 1.1):
51
 
52
- | Slice | pure JSON | parseable | malformed |
53
- |---|---:|---:|---:|
54
- | seed-14, 50 prompts | 88.0% | 88.0% | 12.0% |
55
- | held-out seed-15, 20 prompts | 100.0% | 100.0% | 0.0% |
 
 
 
 
 
 
56
 
57
  ## Looping behavior
58
 
@@ -65,13 +71,19 @@ This is a weight-edited version of [LiquidAI/LFM3-1.7B](https://huggingface.co/L
65
  - Embedding edit: rows 22 (`.`) and 39 (`?`) shifted by +0.10 times the boundary brace-logit direction, where the direction is the boundary state minus the prose and post-brace states.
66
  - All other tensors byte-identical.
67
 
68
- **Final numbers** (standard sampling):
 
 
 
 
69
 
70
  | Slice | ramble rate | median narration words | mean gen tokens |
71
  |---|---:|---:|---:|
72
  | seed-14, 50 prompts | 64.0% | 508 | 1951 |
73
  | held-out seed-15, 20 prompts | 75.0% | 524 | 1848 |
74
 
 
 
75
  ## Files and reproducibility
76
 
77
  - `model.safetensors` - the edited weights (bf16, 3.98 GB).
 
47
  - All other tensors byte-identical; the readout is linear, so each push is exact.
48
  - Mechanism: at antml decision states the newline token (row 207) wins by ~29 logits over both `{` and `<`. Suppressing the newline family along the antml-state direction releases the brace gate.
49
 
50
+ **Slices:** seed 11 is the original held-out slice used to measure the JSON-format edit (50 prompts x 3 samples = 150 per arm). Seed 14 is the development slice used to fit and screen the later edits (50 prompts x 1 sample). Held-out seed 15 is a disjoint slice (20 prompts) the edits never saw, used as a generalization check.
51
 
52
+ **Original baseline (LiquidAI/LFM3-1.7B, seed-11 slice):** pure JSON 42.0%, parseable 68.7%, mean gen tokens 2008.
53
+
54
+ **Final numbers:**
55
+
56
+ | Slice | pure JSON | parseable | malformed | mean gen tokens |
57
+ |---|---:|---:|---:|---:|
58
+ | seed-14, 50 prompts | 88.0% | 88.0% | 12.0% | 1951 |
59
+ | held-out seed-15, 20 prompts | 100.0% | 100.0% | 0.0% | 1848 |
60
+
61
+ **Deltas vs original baseline:** seed-14: +46.0 pts pure JSON, +19.3 pts parseable, -2.8% tokens. Held-out seed-15: +58.0 pts pure JSON, +31.3 pts parseable, -8.0% tokens.
62
 
63
  ## Looping behavior
64
 
 
71
  - Embedding edit: rows 22 (`.`) and 39 (`?`) shifted by +0.10 times the boundary brace-logit direction, where the direction is the boundary state minus the prose and post-brace states.
72
  - All other tensors byte-identical.
73
 
74
+ **Slices:** seed 14 is the development slice used to fit and screen the looping and embedding edits (50 prompts x 1 sample). Held-out seed 15 is a disjoint slice (20 prompts) the edits never saw.
75
+
76
+ **Original baseline (the JSON-format edit, seed-14 slice):** ramble rate 72.0%, mean gen tokens 2238.
77
+
78
+ **Final numbers:**
79
 
80
  | Slice | ramble rate | median narration words | mean gen tokens |
81
  |---|---:|---:|---:|
82
  | seed-14, 50 prompts | 64.0% | 508 | 1951 |
83
  | held-out seed-15, 20 prompts | 75.0% | 524 | 1848 |
84
 
85
+ **Deltas:** seed-14 vs baseline: -8.0 pts ramble, -12.8% tokens. Held-out seed-15 vs the prior version (70.0% ramble, 1514 tokens): +5.0 pts ramble, +22.1% tokens.
86
+
87
  ## Files and reproducibility
88
 
89
  - `model.safetensors` - the edited weights (bf16, 3.98 GB).