yafitzdev commited on
Commit
55a0162
·
verified ·
1 Parent(s): 514a9dc

Update Pyrrho v2 model card terminology

Browse files
Files changed (1) hide show
  1. README.md +5 -11
README.md CHANGED
@@ -14,8 +14,8 @@ tags:
14
 
15
  # pyrrho-v2-nano-g1
16
 
17
- `pyrrho-v2-nano-g1` is a ModernBERT-base classifier for Fitz RAG governance.
18
- It replaces the older Pyrrho g5 multitask shape with four v2 heads:
19
 
20
  - `evidence_verdict`: `INSUFFICIENT`, `DISPUTED`, `SUFFICIENT`
21
  - `failure_mode`: `none`, `unresolved_conflict`, `missing_or_incomplete_evidence`, `wrong_scope_or_version`, `ambiguous_request`
@@ -24,13 +24,11 @@ It replaces the older Pyrrho g5 multitask shape with four v2 heads:
24
 
25
  The model is intended for local governance in `fitz-sage`: decide whether
26
  retrieved evidence is sufficient, insufficient, or disputed, and expose
27
- actionable retrieval/failure metadata. In `fitz-sage`, v2 g1 is used for
28
- post-retrieval evidence governance. Query-only pre-retrieval planning remains
29
- disabled until a query-trained v2 head is available.
30
 
31
  ## Inputs
32
 
33
- Post-retrieval governance input:
34
 
35
  ```text
36
  Question: <user query>
@@ -40,9 +38,6 @@ Sources:
40
  [2] <retrieved source text>
41
  ```
42
 
43
- The raw model can be run on query-only text, but `fitz-sage` does not use those
44
- logits for first-pass retrieval planning in this release.
45
-
46
  ## Output Decoding
47
 
48
  The 18 logits are not one flat softmax. Decode them by group:
@@ -100,7 +95,6 @@ kinds = [
100
  - Dataset: `fitz-gov-v2`
101
  - Clean active training rows: 41,358
102
  - Training source pointer: `fitz_gov_v2_41358_20260703`
103
- - Poisoned/quarantined later data is excluded.
104
  - Base model: `answerdotai/ModernBERT-base`
105
  - Seed: 42
106
 
@@ -123,7 +117,7 @@ Fitz-sage benchmark check for this release candidate:
123
 
124
  | Benchmark | Result |
125
  | --- | ---: |
126
- | balanced fixed-evidence toy cases | 120/120 |
127
  | live fitz-sage benchmark | 86/120 |
128
 
129
  The live benchmark result is the practical integration target; the fixed-evidence
 
14
 
15
  # pyrrho-v2-nano-g1
16
 
17
+ `pyrrho-v2-nano-g1` is the Pyrrho v2 ModernBERT-base classifier for Fitz RAG
18
+ governance. It emits four native v2 heads:
19
 
20
  - `evidence_verdict`: `INSUFFICIENT`, `DISPUTED`, `SUFFICIENT`
21
  - `failure_mode`: `none`, `unresolved_conflict`, `missing_or_incomplete_evidence`, `wrong_scope_or_version`, `ambiguous_request`
 
24
 
25
  The model is intended for local governance in `fitz-sage`: decide whether
26
  retrieved evidence is sufficient, insufficient, or disputed, and expose
27
+ actionable retrieval/failure metadata.
 
 
28
 
29
  ## Inputs
30
 
31
+ Governance input:
32
 
33
  ```text
34
  Question: <user query>
 
38
  [2] <retrieved source text>
39
  ```
40
 
 
 
 
41
  ## Output Decoding
42
 
43
  The 18 logits are not one flat softmax. Decode them by group:
 
95
  - Dataset: `fitz-gov-v2`
96
  - Clean active training rows: 41,358
97
  - Training source pointer: `fitz_gov_v2_41358_20260703`
 
98
  - Base model: `answerdotai/ModernBERT-base`
99
  - Seed: 42
100
 
 
117
 
118
  | Benchmark | Result |
119
  | --- | ---: |
120
+ | balanced fixed-evidence governance sanity suite | 120/120 |
121
  | live fitz-sage benchmark | 86/120 |
122
 
123
  The live benchmark result is the practical integration target; the fixed-evidence