consolidate: drop the duplicated root copy
Browse files- README.md +0 -43
- config.json +0 -33
- generation_config.json +0 -6
- model.safetensors +0 -3
- tokenizer.json +0 -0
- tokenizer_config.json +0 -15
README.md
DELETED
|
@@ -1,43 +0,0 @@
|
|
| 1 |
-
---
|
| 2 |
-
library_name: transformers
|
| 3 |
-
tags: [model-merging, mergeability, training-free, quotient-merge-distance]
|
| 4 |
-
---
|
| 5 |
-
|
| 6 |
-
# pythia-en-zh-14m__average__naive
|
| 7 |
-
|
| 8 |
-
Training-free merged checkpoint from the **Mergeability** sweep
|
| 9 |
-
(`benchmark/emit_lm.py --real`), produced by weight-space merging of two independently
|
| 10 |
-
trained parents. No gradient steps were taken.
|
| 11 |
-
|
| 12 |
-
| field | value |
|
| 13 |
-
|---|---|
|
| 14 |
-
| pair_id | `pythia-en-zh-14m` |
|
| 15 |
-
| parent_a | `EleutherAI/pythia-14m` |
|
| 16 |
-
| parent_b | `SJTU-CL/Zh-Pythia-14M` |
|
| 17 |
-
| ceiling | `(floor-relative)` |
|
| 18 |
-
| operator | `average` |
|
| 19 |
-
| alignment | `naive` |
|
| 20 |
-
| align_method | `permutation` |
|
| 21 |
-
| regime | `different_corpora` |
|
| 22 |
-
| eval_langs | `eng+zho` |
|
| 23 |
-
| nll_merge | `12.0439` |
|
| 24 |
-
| nll_floor | `4.8059` |
|
| 25 |
-
| param_coverage | `0.0846` |
|
| 26 |
-
| MS | `None` |
|
| 27 |
-
|
| 28 |
-
## How it was made
|
| 29 |
-
|
| 30 |
-
Parents were loaded, activations extracted on a shared calibration corpus, and the merge applied
|
| 31 |
-
either **naive** (parents combined in their own coordinates) or **aligned** (parent B carried into
|
| 32 |
-
parent A's residual-stream basis via `common.alignment.residual_basis_map` before merging —
|
| 33 |
-
permutation for same-width pairs, orthogonal/rectangular for cross-width).
|
| 34 |
-
|
| 35 |
-
`MS` is the recovery score from `common.eval.mergeability_score` (merged vs. floor vs. ceiling), the
|
| 36 |
-
same normalisation used by Zhou et al., so it is comparable across rows of the sweep.
|
| 37 |
-
|
| 38 |
-
## Caveats
|
| 39 |
-
|
| 40 |
-
Sub-1B merges are noisy; an aligned signal where the naive one is noise is the finding, not a bug.
|
| 41 |
-
Rows without a joint ceiling are floor-relative and must not be read as absolute recovery.
|
| 42 |
-
|
| 43 |
-
Generated automatically — see the [mergeability repo](https://github.com/suchirsalhan/mergeability).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
config.json
DELETED
|
@@ -1,33 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"architectures": [
|
| 3 |
-
"GPTNeoXForCausalLM"
|
| 4 |
-
],
|
| 5 |
-
"attention_bias": true,
|
| 6 |
-
"attention_dropout": 0.0,
|
| 7 |
-
"bos_token_id": 0,
|
| 8 |
-
"classifier_dropout": 0.1,
|
| 9 |
-
"dtype": "float16",
|
| 10 |
-
"eos_token_id": 0,
|
| 11 |
-
"hidden_act": "gelu",
|
| 12 |
-
"hidden_dropout": 0.0,
|
| 13 |
-
"hidden_size": 128,
|
| 14 |
-
"initializer_range": 0.02,
|
| 15 |
-
"intermediate_size": 512,
|
| 16 |
-
"is_decoder": false,
|
| 17 |
-
"layer_norm_eps": 1e-05,
|
| 18 |
-
"max_position_embeddings": 2048,
|
| 19 |
-
"model_type": "gpt_neox",
|
| 20 |
-
"num_attention_heads": 4,
|
| 21 |
-
"num_hidden_layers": 6,
|
| 22 |
-
"pad_token_id": null,
|
| 23 |
-
"rope_parameters": {
|
| 24 |
-
"partial_rotary_factor": 0.25,
|
| 25 |
-
"rope_theta": 10000,
|
| 26 |
-
"rope_type": "default"
|
| 27 |
-
},
|
| 28 |
-
"tie_word_embeddings": false,
|
| 29 |
-
"transformers_version": "5.14.1",
|
| 30 |
-
"use_cache": true,
|
| 31 |
-
"use_parallel_residual": true,
|
| 32 |
-
"vocab_size": 50304
|
| 33 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
generation_config.json
DELETED
|
@@ -1,6 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"_from_model_config": true,
|
| 3 |
-
"bos_token_id": 0,
|
| 4 |
-
"eos_token_id": 0,
|
| 5 |
-
"transformers_version": "5.14.1"
|
| 6 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
model.safetensors
DELETED
|
@@ -1,3 +0,0 @@
|
|
| 1 |
-
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:b4a18f3c63ec22c2258c173cf47051382b42c9f9edac3bedd08081a41f40ba9f
|
| 3 |
-
size 28143920
|
|
|
|
|
|
|
|
|
|
|
|
tokenizer.json
DELETED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
DELETED
|
@@ -1,15 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"add_prefix_space": false,
|
| 3 |
-
"backend": "tokenizers",
|
| 4 |
-
"bos_token": "<|endoftext|>",
|
| 5 |
-
"clean_up_tokenization_spaces": true,
|
| 6 |
-
"eos_token": "<|endoftext|>",
|
| 7 |
-
"errors": "replace",
|
| 8 |
-
"is_local": false,
|
| 9 |
-
"local_files_only": false,
|
| 10 |
-
"model_max_length": 1000000000000000019884624838656,
|
| 11 |
-
"pad_token": "<|endoftext|>",
|
| 12 |
-
"tokenizer_class": "GPTNeoXTokenizer",
|
| 13 |
-
"trim_offsets": true,
|
| 14 |
-
"unk_token": "<|endoftext|>"
|
| 15 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|