Views
No views yet
| Parameter | Value |
|---|---|
| direction_index | 25.53 |
| attn.o_proj.max_weight | 1.50 |
| attn.o_proj.max_weight_position | 32.73 |
| attn.o_proj.min_weight | 1.43 |
| attn.o_proj.min_weight_distance | 30.47 |
| mlp.down_proj.max_weight | 0.98 |
| mlp.down_proj.max_weight_position | 39.99 |
| mlp.down_proj.min_weight | 0.81 |
| mlp.down_proj.min_weight_distance | 16.95 |
| Metric | This model | Original model (TheDrummer/Skyfall-31B-v4.1) |
|---|---|---|
| KL divergence | 0.0134 | 0 (by definition) |
| Refusals | 8/100 | 87/100 |
V4r is really amazing. Even at >16000 tokens on q4, I can introduce new characters and twists and it goes along just fine. Super complex characters with many conflicting emotions and thoughts. My new fave, and I haven't even tried thinking mode yet.
Surprisingly super coherent logically at 16k+ context. At least as a writing assistant. Can see this one be a daily driver on budget. Even without thinking.
It's kinda fantastic. 😍
Amazing model, the best of this series and generally the best that I can run on my hardware. It also works well with Vector Storage in SillyTavern - it brings up already known information more often.
Ran what I like to call the gauntlet. 7 characters with ~800 tkn ctx each (on av, some art more). AI had to balance switching users mind set (each character is linked to dif part of user. using collective consciousness as the logic binder) Char must retain their traits on top of adapting to new information. Three had a different interpretation, but all seven worked together. my poor computer cant handle beyond 16k in it though, far too long PP time as its timing out. Logic was sound, prose was reflective of each character personality (although there was a slight fact mixup that another char ended up ref. had to clean that out). Cohesive story throughout.
Like this one so far. I like how the reasoning is in character too. Go to 32k context without many repeats. I like it. 4-5 chars are not problem either.
config-v4r