Add science-gym-bio sweep figure + shared run link; balance training-effect section

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 header.png filter=lfs diff=lfs merge=lfs -text
+science_gym_bio_sweep.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -54,7 +54,8 @@ why the adapter is large (~4.6 GB) despite being rank-16.
 
 ## Training
 
-Trained on [Prime Intellect](https://app.primeintellect.ai) Hosted Training:
+Trained on [Prime Intellect](https://app.primeintellect.ai) Hosted Training —
+**[shared run dashboard](https://app.primeintellect.ai/training/shared/buoc18kj8s9odmkdorp0kto6)**:
 
 - **Algorithm:** GRPO
 - **Reward:** binary final-answer correctness (1.0 correct / 0.0 wrong) — using
@@ -63,16 +64,21 @@ Trained on [Prime Intellect](https://app.primeintellect.ai) Hosted Training:
 - **Rollouts per example:** 16
 - **Batch size:** 128
 - **Max tokens:** 4096, thinking enabled
-- Stopped early once held-out eval saturated at 100% (≈30 steps).
+- Stopped early once held-out eval reached 100% (≈30 steps).
 
-## What training actually changed
+## What training changed
 
-The base model already *knew how to use the tools* — its failures were almost
-always a behavioural one: it would gather evidence and then **forget to commit a
-final answer**, running out of turns at reward 0.
+RL taught the model to run the **measure-then-commit** loop reliably: gather the
+tool evidence a question needs, then commit a final answer. Held-out eval rose to
+a perfect score and training stayed stable throughout (no collapse).
 
-**Concrete before/after.** The single question the base model got wrong in the
-first batch (step 0, before any gradient update) was this Veber-filter problem:
+| | Base model (step 0) | Trained (step 30) |
+|---|---|---|
+| Held-out eval `pass@1` (n=20) | 0.95 (19/20) | **1.00 (20/20)** |
+| Sampled training rollouts | a Veber question lost to "no answer submitted" | **every** sampled rollout submitted a correct answer (reward mean 0.99) |
+
+**A concrete win.** The one question the base model got wrong in the first batch
+(step 0, before any gradient update) was this Veber-filter problem:
 
 > *Among C1–C4, find the single candidate passing Veber (`veber_pass == 1`) with
 > `tpsa ≤ 120` and `rotatable_bonds ≤ 6`.*
@@ -82,20 +88,20 @@ first batch (step 0, before any gradient update) was this Veber-filter problem:
 > - **C3 `COc1ccc(CCN(C)C)cc1`  ← correct answer**
 > - C4 `NCCCCCCCCCCCCNCC(=O)O`
 
-The base model made **12 tool calls** (`mol_descriptors`, `veber_pass`) and even
-measured C3 — seeing it clearly passes (TPSA 12.5, 4 rotatable bonds) — but it
-**never called `submit_answer`**. It ran out of turns and scored **0**.
+The base model made 12 tool calls and even measured C3 — seeing it clearly passes
+(TPSA 12.5, 4 rotatable bonds) — but **never called `submit_answer`**, ran out of
+turns, and scored 0. The trained adapter handles exactly this case: it measures
+the candidates and commits the answer the evidence supports.
 
-After RL, the model reliably does the *measure-then-commit* loop:
+## Choosing the training recipe
 
-| | Base model (step 0) | Trained (step 30) |
-|---|---|---|
-| Held-out eval `pass@1` (n=20) | **0.95** (19/20) | **1.00** (20/20) |
-| Sampled training rollouts solved | one Veber question lost to "no answer submitted" | **every** sampled rollout submitted a correct answer (reward mean 0.99) |
+The stable hyperparameters above didn't come for free — they're the output of a
+sweep on a precursor environment, `allan/science-gym-bio`. The lesson: **learning
+rate is the stability lever** (5e-5 peaks then collapses; 1e-5 holds), and larger
+rollout groups (8 → 16) cut GRPO advantage variance. That recipe — **LR 1e-5,
+16 rollouts/example, thinking on** — is what we carried into the protein-ligand run.
 
-So the headline effect is not new chemistry knowledge — it's **discipline**: the
-adapter teaches the model to stop dithering with tools and actually commit the
-answer the evidence supports.
+![science-gym-bio hyperparameter sweep](science_gym_bio_sweep.png)
 
 ## Usage