ehartford commited on
Commit
07ac6ff
·
verified ·
1 Parent(s): 24402ea

Model card: stand-alone; no FlyGPT references beyond build-script provenance

Browse files
Files changed (1) hide show
  1. README.md +3 -6
README.md CHANGED
@@ -17,8 +17,7 @@ presynaptic neuron's predicted neurotransmitter, rescales the whole matrix to a
17
  projections, per-neuron gains, and a readout around it. Those are good modeling decisions for that project, but
18
  they are baked into its weights, so anyone who wants the connectome for something else has to undo them or go back
19
  to the raw release. This repository is the layer underneath: the **source connectome, unmodified**, so that
20
- fly-llm-hf, [FlyGPT](https://github.com/QuixiAI/FlyGPT), graph-ML work, and simulations can all start from the same
21
- lossless tensors. At a high level the differences are:
22
 
23
  - **Whole nervous system, not a subset.** Every body in the release's connectivity file: brain, optic lobes, and
24
  ventral nerve cord, with subsets available as masks rather than chosen for you.
@@ -67,7 +66,7 @@ Facts about fly-llm-hf are taken from its model card.
67
  | `transformers` role | `AutoModel` returning the graph | `AutoModelForCausalLM` generating text |
68
 
69
  ```text
70
- Janelia MaleCNS v1.0 ──► this repository (synapse counts, untouched) ──► FlyGPT (its own trainable edge weights on the same topology)
71
  Janelia MaleCNS v1.0 ──► Xenova packaging ──► fly-llm-hf (signed, rescaled, frozen reservoir + TinyStories readout)
72
  ```
73
 
@@ -153,7 +152,7 @@ incoming = cns(x) # one propagati
153
 
154
  ## Provenance
155
 
156
- Built by [`data/fly/export_malecns_hf.py`](https://github.com/QuixiAI/FlyGPT) from these files in
157
  `gs://flyem-male-cns/v1.0/connectome-data/flat-connectome/`, md5-verified against the bucket listing:
158
 
159
  | role | file | md5 (base64) |
@@ -211,5 +210,3 @@ license; attribution to the original authors (above) is required.
211
 
212
  Official site: https://male-cns.janelia.org · neuPrint dataset `male-cns:v1.0`.
213
 
214
- This repository is the data backbone of [FlyGPT](https://github.com/QuixiAI/FlyGPT), which trains a language model
215
- whose recurrent architecture is a subgraph of this connectome, with its own trainable edge weights.
 
17
  projections, per-neuron gains, and a readout around it. Those are good modeling decisions for that project, but
18
  they are baked into its weights, so anyone who wants the connectome for something else has to undo them or go back
19
  to the raw release. This repository is the layer underneath: the **source connectome, unmodified**, so that
20
+ fly-llm-hf, downstream models, graph-ML work, and simulations can all start from the same lossless tensors. At a high level the differences are:
 
21
 
22
  - **Whole nervous system, not a subset.** Every body in the release's connectivity file: brain, optic lobes, and
23
  ventral nerve cord, with subsets available as masks rather than chosen for you.
 
66
  | `transformers` role | `AutoModel` returning the graph | `AutoModelForCausalLM` generating text |
67
 
68
  ```text
69
+ Janelia MaleCNS v1.0 ──► this repository (synapse counts, untouched) ──► your model (whatever weights you derive from them)
70
  Janelia MaleCNS v1.0 ──► Xenova packaging ──► fly-llm-hf (signed, rescaled, frozen reservoir + TinyStories readout)
71
  ```
72
 
 
152
 
153
  ## Provenance
154
 
155
+ Built by the script [`data/fly/export_malecns_hf.py`](https://github.com/QuixiAI/FlyGPT/blob/main/data/fly/export_malecns_hf.py) from these files in
156
  `gs://flyem-male-cns/v1.0/connectome-data/flat-connectome/`, md5-verified against the bucket listing:
157
 
158
  | role | file | md5 (base64) |
 
210
 
211
  Official site: https://male-cns.janelia.org · neuPrint dataset `male-cns:v1.0`.
212