Add model card with checkpoint table
Browse files
README.md
CHANGED
|
@@ -1,3 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# Latent CoT curriculum + backtracking (2-arm star)
|
| 2 |
|
| 3 |
Fork / extension of [Reasoning by Superposition](https://arxiv.org/abs/2505.12514) ([original repo](https://github.com/Ber666/reasoning-by-superposition)).
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
library_name: pytorch
|
| 4 |
+
tags:
|
| 5 |
+
- coconut
|
| 6 |
+
- latent-cot
|
| 7 |
+
- graph-reasoning
|
| 8 |
+
- backtracking
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
Code + weights for latent CoT curriculum + backtracking on 2-arm star graphs.
|
| 12 |
+
|
| 13 |
+
**GitHub:** https://github.com/Avra98/reasoning-by-superposition-latent
|
| 14 |
+
|
| 15 |
+
## Shipped checkpoints (raw Coconut `state_dict`, ~59MB each)
|
| 16 |
+
|
| 17 |
+
Download with `hf download Avra98/latent_backtrack <path>`.
|
| 18 |
+
|
| 19 |
+
| Path | What |
|
| 20 |
+
|------|------|
|
| 21 |
+
| `ckpts/star-coconut-L10-bfs-stage0/checkpoint_150` | L10 stage-0 (known-good hop-1) |
|
| 22 |
+
| `ckpts/star-coconut-L15-bfs-stage0-warm/checkpoint_150` | L15 stage-0, warm from L10 |
|
| 23 |
+
| `ckpts/L20_w2_s1_prom090_bt090/checkpoint_225` | L20 BT **W=2** (intervention ckpt) |
|
| 24 |
+
| `ckpts/L20_w2_s1_prom090_bt090/checkpoint_325` | L20 BT W=2 latest |
|
| 25 |
+
| `ckpts/L20_w5_s1_prom090_bt090/checkpoint_225` | L20 BT **W=5** |
|
| 26 |
+
| `ckpts/L20_w5_s1_prom090_bt090/checkpoint_290` | L20 BT W=5 latest |
|
| 27 |
+
| `ckpts/L20_w20_s1_prom090_bt090/checkpoint_345` | L20 **full BPTT** |
|
| 28 |
+
| `ckpts/L20_cso_prom090/checkpoint_200` | L20 CSO baseline |
|
| 29 |
+
|
| 30 |
+
L20 W=2/W=5/full all warm-start from the L15 stage-0 file above (`init_stage: 1`).
|
| 31 |
+
|
| 32 |
+
Load with this repo's `Coconut` wrapper (`run.py` `load_model_path` or `scripts/attention_atlas.py`), not `from_pretrained`.
|
| 33 |
+
|
| 34 |
# Latent CoT curriculum + backtracking (2-arm star)
|
| 35 |
|
| 36 |
Fork / extension of [Reasoning by Superposition](https://arxiv.org/abs/2505.12514) ([original repo](https://github.com/Ber666/reasoning-by-superposition)).
|