rgup commited on
Commit
e34084a
·
verified ·
1 Parent(s): 75b2199

Release package

Browse files
Files changed (1) hide show
  1. README.md +2 -2
README.md CHANGED
@@ -30,13 +30,13 @@ Both builds vs the bf16 model in Mirai's in-house agent circuits (SGLang, Flint'
30
  | Hard-problem battery | 45/55 | 50/55 | 49/55 |
31
  | AppWorld | 168/240 | 157/240 | 187/240 |
32
  | Assistant week v2 (native tools) | 409/656 | 400/656 | 460/656 |
33
- | Assistant week v1 (text markers) | 42/136 | 24/136 | 60/136 |
34
  | Creative build, fully correct | 4/4 | 3/4 | 4/4 |
35
  | Agentic build, declared done | 4/4 | 4/4 | 4/4 |
36
 
37
  Mean answer length is 1,605 tokens for Flint S 11.8 (bf16: 1,623; 99.5% of answers finished within 8k tokens) and 1,710 for Flint S 12.4 (+5%; 98.5% finished).
38
 
39
- The 11.8 GB build ships the weights after 12M tokens of QAT. Training it for 171M tokens lowered its KL to 0.1138 but lost on the agent circuits (paired: assistant week v1 −6.6 points, v2 −5.2 points, 95% intervals excluding zero), so the circuits decide. The 12.4 GB build is the 171M-token checkpoint.
40
 
41
  ## What is in this repository
42
  **These are dense bf16 checkpoints (about 80 GB per build), not the packed 12.4 / 11.8 GB files.** Each folder holds the exact dequantized S weights in the original Flint layout, so they run with Flint's SGLang serving kit (`<build>/sglang/`, one H200 or B200/B300 GPU, as in the base model card) and reproduce the scores above. No runtime reads a packed Flint S package yet.
 
30
  | Hard-problem battery | 45/55 | 50/55 | 49/55 |
31
  | AppWorld | 168/240 | 157/240 | 187/240 |
32
  | Assistant week v2 (native tools) | 409/656 | 400/656 | 460/656 |
33
+ | Assistant week v1 (text markers, 8 seeds) | 72/272 | 65/272 | 60/136 (4 seeds) |
34
  | Creative build, fully correct | 4/4 | 3/4 | 4/4 |
35
  | Agentic build, declared done | 4/4 | 4/4 | 4/4 |
36
 
37
  Mean answer length is 1,605 tokens for Flint S 11.8 (bf16: 1,623; 99.5% of answers finished within 8k tokens) and 1,710 for Flint S 12.4 (+5%; 98.5% finished).
38
 
39
+ The 11.8 GB build ships the weights after 12M tokens of QAT. Training it for 171M tokens lowered its KL to 0.1138 but lost on assistant week v2 (paired, 16 seeds: −5.2 points, 95% interval −9.1 to −1.1; v1 tied at 8 seeds), so the circuits decide. The 12.4 GB build is the 171M-token checkpoint.
40
 
41
  ## What is in this repository
42
  **These are dense bf16 checkpoints (about 80 GB per build), not the packed 12.4 / 11.8 GB files.** Each folder holds the exact dequantized S weights in the original Flint layout, so they run with Flint's SGLang serving kit (`<build>/sglang/`, one H200 or B200/B300 GPU, as in the base model card) and reproduce the scores above. No runtime reads a packed Flint S package yet.