Avifenesh commited on
Commit
f9db98e
·
verified ·
1 Parent(s): 5f6e75e

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +14 -0
README.md CHANGED
@@ -50,6 +50,20 @@ masked head — see
50
  model's *own generations* (163k-token corpus over real agentic session
51
  prompts; external text used as prompts only — the ranked distribution is
52
  100% model-generated). Ranks artifact: `q38-ranks-sxc32768.gguf(.txt)`.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
53
  - **Source model:** ranks generated by, and head extracted from, THIS repo's
54
  `Qwen3.8-27B-NVFP4-Q5K-mtp.gguf` (minted from
55
  [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) via the
 
50
  model's *own generations* (163k-token corpus over real agentic session
51
  prompts; external text used as prompts only — the ranked distribution is
52
  100% model-generated). Ranks artifact: `q38-ranks-sxc32768.gguf(.txt)`.
53
+ - **Flavors (2026-08-15):** three ranks corpora, pick by workload — each ships as a
54
+ ranks list (`.txt`, drives memra's load-time trim on safetensors trunks via
55
+ `MEMRA_FRSPEC_TRIM`) and as a pre-trimmed draft head (`.gguf`, serve with
56
+ `MEMRA_MTP_DRAFT`):
57
+ | flavor | files | corpus |
58
+ |---|---|---|
59
+ | **agentic** (serving default) | `q38-ranks-sxc32768.gguf(.txt)` + `mtp-...frspec-sxc32768.gguf` | 163k own-generated tokens over real agentic sessions |
60
+ | **prose** | `q38-ranks-prose-32768.{gguf,txt}` + `mtp-...frspec-prose32768.gguf` | 154k own-generated tokens over essay/story/letter prompts, ~15% non-English |
61
+ | **mixed** | `q38-ranks-mixed-32768.txt` + `mtp-...frspec-mixed32768.gguf` | 50/50 normalized count-blend of both streams (same rank law) |
62
+
63
+ A mask can never change OUTPUT (verification is the target's job on the full
64
+ vocab); it only moves draft acceptance. Short generic probes measure the three
65
+ within noise of each other (acceptance 0.40 prose-text / 0.59-0.62 code); the
66
+ differences live in domain tail tokens.
67
  - **Source model:** ranks generated by, and head extracted from, THIS repo's
68
  `Qwen3.8-27B-NVFP4-Q5K-mtp.gguf` (minted from
69
  [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) via the