Add LLaMA models (#35)

- added llama-7b (92d7a9ebdd2aabc8a4df00064c7ff20d3d31491f)
- New float16 llama 7b lens (b7ab2257bedd17b296386e7027b78abcbb2bcdb6)
- Add llama 13b lens (95f31c6c3e61996b5f871095c3e6922969d58a89)


Co-authored-by: Nora Belrose <norabelrose@users.noreply.huggingface.co>

.gitattributes

@@ -32,3 +32,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.md filter=lfs diff=lfs merge=lfs -text
+*.pyc filter=lfs diff=lfs merge=lfs -text
+*.py filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text
+*.txt filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,12 +1,3 @@
----
-title: Tuned Lens
-emoji: 🔎
-colorFrom: pink
-colorTo: blue
-sdk: gradio
-python_version: 3.10.2
-sdk_version: 3.20.0
-app_file: app.py
-pinned: false
-license: mit
----
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd20cba45293ebf4604bfc781e4ebd35a46ce45ff0c9c161492c6fb9f4912d5b
+size 167

app.py

@@ -1,112 +1,3 @@
-import torch
-from tuned_lens.nn.lenses import TunedLens, LogitLens
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from tuned_lens.plotting import plot_lens
-import gradio as gr
-from plotly import graph_objects as go
-
-device = torch.device("cpu")
-print(f"Using device {device} for inference")
-model = AutoModelForCausalLM.from_pretrained("EleutherAI/pythia-410m-deduped-v0")
-model = model.to(device)
-tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pythia-410m-deduped-v0")
-tuned_lens = TunedLens.load("pythia-410m-deduped-v0", map_location=device)
-logit_lens = LogitLens(model)
-
-lens_options_dict = {
-    "Tuned Lens": tuned_lens,
-    "Logit Lens": logit_lens,
-}
-
-statistic_options_dict = {
-    "Entropy": "entropy",
-    "Cross Entropy": "ce",
-    "Forward KL": "forward_kl",
-}
-
-
-def make_plot(lens, text, statistic, token_cutoff):
-    input_ids = tokenizer.encode(text, return_tensors="pt")
-
-    if len(input_ids[0]) == 0:
-        return go.Figure(layout=dict(title="Please enter some text."))
-
-    if token_cutoff < 1:
-        return go.Figure(layout=dict(title="Please provide valid token cut off."))
-
-    fig = plot_lens(
-        model,
-        tokenizer,
-        lens_options_dict[lens],
-        layer_stride=2,
-        input_ids=input_ids,
-        start_pos=max(len(input_ids[0]) - token_cutoff, 0),
-        statistic=statistic_options_dict[statistic],
-    )
-
-    return fig
-
-
-preamble = """
-# The Tuned Lens 🔎
-
-A tuned lens allows us to peak at the iterative computations a transformer uses to compute the next token.
-
-A lens into a transformer with n layers allows you to replace the last $m$ layers of the model with an [affine transformation](https://pytorch.org/docs/stable/generated/torch.nn.Linear.html) (we call these affine translators).
-
-This essentially skips over these last few layers and lets you see the best prediction that can be made from the model's representations, i.e. the residual stream, at layer $n - m$. Since the representations may be rotated, shifted, or stretched from layer to layer it's useful to train the len's affine adapters specifically on each layer. This training is what differentiates this method from simpler approaches that decode the residual stream of the network directly using the unembeding layer i.e. the logit lens. We explain this process in [the paper](https://arxiv.org/abs/2303.08112).
-
-## Usage
-Since the tuned lens produces a distribution of predictions to visualize it's output we need to we need to provide a summary statistic to plot.  The default is simply [entropy](https://en.wikipedia.org/wiki/Entropy_(information_theory)), but you can also choose the [cross entropy](https://en.wikipedia.org/wiki/Cross_entropy) with the target token, or the [KL divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence) between the model's predictions and the tuned lens' predictions. You can also hover over a token to see more of the distribution i.e. the top 10 most probable tokens and their probabilities.
-
-## Examples
-Here are some interesting examples you can try.
-
-### Copy paste:
-```
-Copy: A!2j!#u&NGApS&MkkHe8Gm!#
-Paste: A!2j!#u&NGApS&MkkHe8Gm!#
-```
-
-### Trivial in-context learning
-```
-inc 1 2
-inc 4 5
-inc 13 
-```
-
-#### Addition
-```
-add 1 1 2
-add 3 4 7
-add 13 2 
-```
-"""
-
-with gr.Blocks() as demo:
-    gr.Markdown(preamble)
-    with gr.Column():
-        text = gr.Textbox(
-            value="it was the best of times, it was the worst of times",
-            label="Input Text",
-        )
-        with gr.Row():
-            lens_options = gr.Dropdown(
-                list(lens_options_dict.keys()), value="Tuned Lens", label="Select Lens"
-            )
-            statistic = gr.Dropdown(
-                list(statistic_options_dict.keys()),
-                value="Entropy",
-                label="Select Statistic",
-            )
-            token_cutoff = gr.Slider(
-                maximum=20, minimum=2, value=10, step=1, label="Plot Last N Tokens"
-            )
-        examine_btn = gr.Button(value="Submit")
-        plot = gr.Plot()
-    examine_btn.click(make_plot, [lens_options, text, statistic, token_cutoff], plot)
-    demo.load(make_plot, [lens_options, text, statistic, token_cutoff], plot)
-
-if __name__ == "__main__":
-    demo.launch()
-
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc89bb5f80d8a86f0d65fba048f8b4d35e099f03ad268a08e2ccbf6eac8cffbc
+size 4260

lens/EleutherAI/gpt-neox-20b/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/gpt-neox-20b", "d_model": 6144, "num_hidden_layers": 44, "bias": true, "base_model_revision": "4e49eadb5d14bd22f314ec3f45b69a87b88c7691", "unemebd_hash": "323d4c731c33556e143503e3be913c109ead330080b4065552be97000c19ed67", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:f42ec7fc2a0d16d2fdb6ac2f1361d9e775179d41168a7bb12026433ea1ceb168
+size 295

lens/EleutherAI/pythia-1.4b-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-1.4b-deduped-v0", "d_model": 2048, "num_hidden_layers": 24, "bias": true, "base_model_revision": "b541e01fddacd3038799915cf8ff5b52e835a6c4", "unemebd_hash": "da1780eccec1a4ff12e43464da6cbef33b9ffde398a3056ac9648dd53229943e", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:e697309953fc2c15a215c381ea22d91d90cddf739df6ddd54fa8f01c45248344
+size 305

lens/EleutherAI/pythia-1.4b-deduped/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-1.4b-deduped", "d_model": 2048, "num_hidden_layers": 24, "bias": true, "base_model_revision": null, "unemebd_hash": "42dcb52bd3d955648e9fbae0ab40527eed42b22b06945c680255d620d46dffd8", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:c496b7bc74699436b65791838cb1fdf48d27744cb243874af9da25b9be84b465
+size 264

lens/EleutherAI/pythia-12b-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-12b-deduped-v0", "d_model": 5120, "num_hidden_layers": 36, "bias": true, "base_model_revision": "b497662035bf3c80b4f6a1ddfe09bc27763e843a", "unemebd_hash": "a161c0d1dd8793ca1683b0422f3b6573178ea7ebf26cf207e40cc56507aa0526", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f5c244a5d4c5ac6c5bb22d42c922eac03680625605400db384d3a3da1875206
+size 304

lens/EleutherAI/pythia-12b-deduped/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-12b-deduped", "d_model": 5120, "num_hidden_layers": 36, "bias": true, "base_model_revision": null, "unemebd_hash": "d676a152f789a57219afa5b93c3718bf56805a8fcb45eeaacdb6665aa880e82e", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:227869e1813ba53c131cbfbe0cb82d3f556e07cfb704606bb94e846fe6b5f8c5
+size 263

lens/EleutherAI/pythia-160m-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-160m-deduped-v0", "d_model": 768, "num_hidden_layers": 12, "bias": true, "base_model_revision": "7e57cc978f5da949f028f36b5baf8f5d6c3281b1", "unemebd_hash": "922e5aee39d4874fb5c1163087858333808367bf9b02c4a1ae4a06828af2f58a", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2cfc2289ff688df53f599cda06dcac9f72087efe42da8f20a3fda203341413d
+size 304

lens/EleutherAI/pythia-160m-deduped/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-160m-deduped", "d_model": 768, "num_hidden_layers": 12, "bias": true, "base_model_revision": null, "unemebd_hash": "7f56aa8e6e19b2e0dc2d465ca802f839d691a9306a6bca06ccae0cbd17e03283", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:0393bfdd60f616c9606b1db801332c5e1cc1abbf7a52f0b8fc362547d613691c
+size 263

lens/EleutherAI/pythia-1b-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-1b-deduped-v0", "d_model": 2048, "num_hidden_layers": 16, "bias": true, "base_model_revision": "021f79f50ff000ae1c159e22402ffec62284664d", "unemebd_hash": "b97dd35a220ea2694e263be05d2f19129fc5725c1d201d83eae5a78eeebcf527", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc83a156780cb5d49778dfbae742c5c2e1af36123f7cf60764f1adae80ede4bb
+size 303

lens/EleutherAI/pythia-2.8b-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-2.8b-deduped-v0", "d_model": 2560, "num_hidden_layers": 32, "bias": true, "base_model_revision": null, "unemebd_hash": "441a92c8cdb05202cc8f9115039d3747aed8363540d51f7f60a90dbac50e9680", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:06238f3af11e14f5663b498d82e6da7dca87c9e90877817f474fee5fda1f1c67
+size 267

lens/EleutherAI/pythia-2.8b-deduped/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-2.8b-deduped", "d_model": 2560, "num_hidden_layers": 32, "bias": true, "base_model_revision": null, "unemebd_hash": "0e9d6241f727a812fdd9e612496192e1da3409fc0a9426a1077eca654180c6f1", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab43342448b067918933af8e082c6b6a2b4111a80e955bb75577a81f9bec8d4c
+size 264

lens/EleutherAI/pythia-410m-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-410m-deduped-v0", "d_model": 1024, "num_hidden_layers": 24, "bias": true, "base_model_revision": "3538d3569a7e313e445ad6401c92c6e16777a2da", "unemebd_hash": "281af3dac813ef2f2eb5a1a359c402627bc9cf104710d00f891f767b17687758", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:33a2e63741f52966e5ff78d81a91e3d103dbe979d3a03165bedcf071e7e10358
+size 305

lens/EleutherAI/pythia-410m-deduped/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-410m-deduped", "d_model": 1024, "num_hidden_layers": 24, "bias": true, "base_model_revision": null, "unemebd_hash": "e9b52de378cf3ce115b2f23028993fb797268389a10a6241fe981e361dfd2b78", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:9785f07bfea0affe284f39780e87347135725391f910a6846d45a7a18d982006
+size 264

lens/EleutherAI/pythia-6.9b-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-6.9b-deduped-v0", "d_model": 4096, "num_hidden_layers": 32, "bias": true, "base_model_revision": "cbd53efc2e56056e3bd0235277b5d0b668a6dfbb", "unemebd_hash": "5a037e6f7542abd5e0817e46c7e9127c18164ac34d06051b5faac190103f6951", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:b58ab18fa92a144b18fcdb0cb6d24f1d9fe2c2a90bbd69cfab9688e96bcde5a0
+size 305

lens/EleutherAI/pythia-6.9b-deduped/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-6.9b-deduped", "d_model": 4096, "num_hidden_layers": 32, "bias": true, "base_model_revision": null, "unemebd_hash": "272e395758b795a73dddd10c3b33000a33a9380784b963623a36a50d5df6c34e", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe20b7d1dfaa0182930130e349204d9144af9cc1ca0b58d8102d6b81f63cd5f9
+size 264

lens/EleutherAI/pythia-70m-deduped-v0/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-70m-deduped-v0", "d_model": 512, "num_hidden_layers": 6, "bias": true, "base_model_revision": "ec30f7539a604fcb0b7fbba04fb1eb0110735d29", "unemebd_hash": "6c42572c654f76afb6ad30aafac2644308d5e3e708ee54051fa9d4e043918f3a", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:468abebefbad4f03c38bebd45317218fa0ee87ac81edcf963ee5c1af60ff218f
+size 302

lens/EleutherAI/pythia-70m-deduped/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "EleutherAI/pythia-70m-deduped", "d_model": 512, "num_hidden_layers": 6, "bias": true, "base_model_revision": null, "unemebd_hash": "35a97f97ca7cc395758d14911297122458e306a960feab2a9b79d07439de4db4", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:3925c5ff7bde19de9a09bf86ee303757cdc3dc70058f0a9b9528ea357ad6afc7
+size 261

lens/facebook/llama-13b/config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6777d20cd9652dc252752c5c776c07e7c8d920afea9b962c841a16056d3a109
+size 254

lens/facebook/llama-13b/params.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2700859b4335f4e4bdd4040232f594170bb717f8af04f65e5560c49cfb6da122
+size 2097581027

lens/facebook/llama-7b/config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6644680efdcc19840bf75c96e3ab97fb1b16a9e22f2b0456e9b05ff9e7c0d60d
+size 253

lens/facebook/llama-7b/params.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02186a08b29e7173ae229218823b65891a7b2f2eadf0e69baf44a9c62b55a7cf
+size 1074019491

lens/facebook/opt-1.3b/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "facebook/opt-1.3b", "d_model": 2048, "num_hidden_layers": 24, "bias": true, "base_model_revision": "8c7b10754972749675d22364c25c428b29face51", "unemebd_hash": "2db68eed8b11e46e8a969c14b1ce9269edec3154b19cdd18970dcfc405533070", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f6cee11c783ee10cf15b10a4a059b838643a414384584a9c53adeaee265de53
+size 289

lens/facebook/opt-125m/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "facebook/opt-125m", "d_model": 768, "num_hidden_layers": 12, "bias": true, "base_model_revision": "3d2b5f275bdf882b8775f902e1bfdb790e2cfc32", "unemebd_hash": "d54b1bdd7e16d4dab3bb9f1856c1146310a3ce228e667640843455a3956dc9b4", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:e653efc95c2587a4086fef70dd33af22fa4f2e49414668559bb68b8e14fb6b94
+size 288

lens/facebook/opt-6.7b/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "facebook/opt-6.7b", "d_model": 4096, "num_hidden_layers": 32, "bias": true, "base_model_revision": "a45aa65bbeb77c1558bc99bedc6779195462dab0", "unemebd_hash": "35676bc5e38da5b53231218f1c829b91bc89de7f65fec1b2fe885b9c42f93dcb", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:37a2047cc1ea31a9bb0e97c6c84cc9726399981ac76ef16e45a7a0d3499ff5a7
+size 289

lens/gpt2-large/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "gpt2-large", "d_model": 1280, "num_hidden_layers": 36, "bias": true, "base_model_revision": "212095d5832abbf9926672e1c1e8d14312a3be20", "unemebd_hash": "9b7da774c0a326716dca888539370ddff25804795949e5ace65ef9f761f47397", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0650af69d4c7d4ac6ad6e14c604b3fe2e296f967d21afe43017cd20160c5abd
+size 282

lens/gpt2-xl/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "gpt2-xl", "d_model": 1600, "num_hidden_layers": 48, "bias": true, "base_model_revision": "33cdb5c0db5423c1879b1b9f16c352988e8754a8", "unemebd_hash": "70bf58a8cf7964b39530e30fdaebb89de39489546244437b1ed56fb81bd4c746", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:d291da4602b3f32e85563e53ae68dc4c378a00f81f3ee52805f38c03636386fa
+size 279

lens/gpt2/config.json

@@ -1 +1,3 @@
-{"base_model_name_or_path": "gpt2", "d_model": 768, "num_hidden_layers": 12, "bias": true, "base_model_revision": "e7da7f221d5bf496a48136c0cd264e630fe9fcc8", "unemebd_hash": "608e50247f57691c90453601e854f2287141e4db9cba436af0b0186003e2daae", "lens_type": "linear_tuned_lens"}
+version https://git-lfs.github.com/spec/v1
+oid sha256:7269c86486bd48001867e8d040f5e8a41f6a889648bd09548e27f4e59dabc35a
+size 275

requirements.txt

@@ -1 +1,3 @@
-tuned_lens==0.0.5
+version https://git-lfs.github.com/spec/v1
+oid sha256:0de533a67fd8d52c57fce9de893dd7c8c53dfb750bc910848950c18f2ebc80ea
+size 18