Fix OLMo-Think training: proper </think> tag parsing

- Add OLMo to thinking model detection list
- Require </think> tag for thinking models (no tag = 0 reward)
- Extract only content AFTER </think> for poem evaluation
- Reduce max_tokens to 2000 (was 3000, caused OOM)
- Removed heuristic extraction code per user direction

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Author: darrenangle

.beads/issues.jsonl

@@ -54,6 +54,7 @@
 {"id":"abide-4k6","title":"Ghazal form implementation","description":"5-15 couplets, AA BA CA pattern with radif (refrain) and qafiya (rhyme)","status":"closed","priority":1,"issue_type":"task","created_at":"2025-12-07T14:29:16.027733238-06:00","updated_at":"2025-12-07T14:44:48.767802115-06:00","closed_at":"2025-12-07T14:44:48.767802115-06:00","dependencies":[{"issue_id":"abide-4k6","depends_on_id":"abide-rgl","type":"blocks","created_at":"2025-12-07T14:29:36.474694476-06:00","created_by":"daemon"}]}
 {"id":"abide-4pq7","title":"Audit StaircasePoem verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:55:22.217891418-06:00","updated_at":"2025-12-11T11:08:01.630787869-06:00","closed_at":"2025-12-11T11:08:01.630787869-06:00"}
 {"id":"abide-4tf","title":"Audit ChantRoyal verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:45:53.232943326-06:00","updated_at":"2025-12-11T11:08:01.639200354-06:00","closed_at":"2025-12-11T11:08:01.639200354-06:00"}
+{"id":"abide-4x3i","title":"Update training script and start background training","description":"Apply optimal params to train_grpo.py and launch training","status":"open","priority":2,"issue_type":"task","created_at":"2025-12-23T20:52:41.086438031-06:00","updated_at":"2025-12-23T20:52:41.086438031-06:00","dependencies":[{"issue_id":"abide-4x3i","depends_on_id":"abide-zuf1","type":"blocks","created_at":"2025-12-23T20:53:13.411609274-06:00","created_by":"darren"}]}
 {"id":"abide-4yd","title":"Audit HeroicQuatrain verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:49:21.071203699-06:00","updated_at":"2025-12-11T11:02:03.354124376-06:00","closed_at":"2025-12-11T11:02:03.354124376-06:00"}
 {"id":"abide-54yb","title":"Audit Pantoum verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:51:30.05648361-06:00","updated_at":"2025-12-11T11:03:19.370024195-06:00","closed_at":"2025-12-11T11:03:19.370024195-06:00"}
 {"id":"abide-571a","title":"Audit scoring: IsolatedCouplet","description":"Check verify() scoring - fix if too lenient","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T07:09:57.93472223-06:00","updated_at":"2025-12-23T07:21:13.441592785-06:00","closed_at":"2025-12-23T07:21:13.441592785-06:00"}
@@ -104,9 +105,11 @@
 {"id":"abide-8qt","title":"Review novel forms for difficulty and coherence","description":"Review all created forms to ensure they are: difficult enough, unusual/atypical, coherent (possible to write), and truly test instruction-following vs memorization.","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-09T06:27:35.708572135-06:00","updated_at":"2025-12-09T06:41:21.433957007-06:00","closed_at":"2025-12-09T06:41:21.433957007-06:00","dependencies":[{"issue_id":"abide-8qt","depends_on_id":"abide-6hp","type":"blocks","created_at":"2025-12-09T06:27:43.221087987-06:00","created_by":"daemon"}]}
 {"id":"abide-8so","title":"Implement new constraint primitives for novel forms","description":"Add constraint types like: Lipogram (exclude letters), ForcedWords (must include specific words), PositionalConstraint (words at specific positions), CharacterCount, WordPattern, etc.","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-09T06:27:33.932871913-06:00","updated_at":"2025-12-09T06:32:54.956461062-06:00","closed_at":"2025-12-09T06:32:54.956461062-06:00","dependencies":[{"issue_id":"abide-8so","depends_on_id":"abide-3il","type":"blocks","created_at":"2025-12-09T06:27:43.195445798-06:00","created_by":"daemon"}]}
 {"id":"abide-8t0f","title":"Audit scoring: Lai","description":"Check verify() scoring - fix if too lenient","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T07:10:18.301686732-06:00","updated_at":"2025-12-23T07:26:16.54134485-06:00","closed_at":"2025-12-23T07:26:16.54134485-06:00"}
+{"id":"abide-8teh","title":"Run parameter search experiments","description":"Run 5-10 experiments with 128 rollouts each, varying max_tokens and repetition_penalty","status":"in_progress","priority":2,"issue_type":"task","created_at":"2025-12-23T20:52:28.631077569-06:00","updated_at":"2025-12-23T20:56:18.520763259-06:00","dependencies":[{"issue_id":"abide-8teh","depends_on_id":"abide-95wn","type":"blocks","created_at":"2025-12-23T20:53:03.26382537-06:00","created_by":"darren"}]}
 {"id":"abide-8vxw","title":"Audit scoring: AlphabeticTerminus","description":"Check verify() scoring - fix if too lenient","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T07:05:13.347628413-06:00","updated_at":"2025-12-23T07:26:20.032665355-06:00","closed_at":"2025-12-23T07:26:20.032665355-06:00"}
 {"id":"abide-8xmd","title":"Audit Rondine verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:53:44.730691131-06:00","updated_at":"2025-12-11T11:05:19.336969253-06:00","closed_at":"2025-12-11T11:05:19.336969253-06:00"}
 {"id":"abide-8z6","title":"Add refrain verification to DoubleBallade","description":"\n## Current State\nDoubleBallade verifies: LineCount(52), StanzaCount(7), StanzaSizes, RhymeScheme, Syllables\n\n## Missing Verification\nDocstring says: 'Each stanza ends with refrain'\nRefrain should appear at end of each of the 6 main stanzas (lines 8, 16, 24, 32, 40, 48)\nPlus possibly in envoi\n\n## Implementation\nAdd Refrain constraint for stanza-ending lines\n0-indexed: reference_line=7, repeat_at=[15, 23, 31, 39, 47]\n\n## Files\n- src/abide/forms/medieval.py:125-218\n","status":"closed","priority":1,"issue_type":"task","created_at":"2025-12-11T00:47:30.304773485-06:00","updated_at":"2025-12-11T00:54:48.346874949-06:00","closed_at":"2025-12-11T00:54:48.346874949-06:00"}
+{"id":"abide-95wn","title":"Create experiment script and log","description":"Set up param_search.py and experiments/grpo_param_search.md","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T20:52:22.476571837-06:00","updated_at":"2025-12-23T20:56:12.654323075-06:00","closed_at":"2025-12-23T20:56:12.654323075-06:00","dependencies":[{"issue_id":"abide-95wn","depends_on_id":"abide-l12q","type":"blocks","created_at":"2025-12-23T20:52:58.173111956-06:00","created_by":"darren"}]}
 {"id":"abide-97j","title":"Implement Meter enum and stress pattern detection","description":"Create MeterType enum (IAMBIC, TROCHAIC, ANAPESTIC, DACTYLIC, SPONDAIC, PYRRHIC) and functions to detect/score meter from CMU stress patterns","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-07T18:45:22.665235626-06:00","updated_at":"2025-12-07T18:52:32.551679529-06:00","closed_at":"2025-12-07T18:52:32.551679529-06:00"}
 {"id":"abide-9adm","title":"Audit QuestionQuest verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:52:51.936982285-06:00","updated_at":"2025-12-11T11:06:27.97065607-06:00","closed_at":"2025-12-11T11:06:27.97065607-06:00"}
 {"id":"abide-9ct","title":"Add instruction() method to constraints for LLM prompt generation","description":"","status":"closed","priority":2,"issue_type":"feature","created_at":"2025-12-07T14:53:10.637844155-06:00","updated_at":"2025-12-07T14:57:45.375198169-06:00","closed_at":"2025-12-07T14:57:45.375198169-06:00"}
@@ -236,6 +239,7 @@
 {"id":"abide-kmt","title":"Audit HeroicCouplet verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:49:15.995965671-06:00","updated_at":"2025-12-11T11:01:20.996137709-06:00","closed_at":"2025-12-11T11:01:20.996137709-06:00"}
 {"id":"abide-krkx","title":"Audit scoring: HourglassVerse","description":"Check verify() scoring - fix if too lenient","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T07:09:47.703775766-06:00","updated_at":"2025-12-23T07:25:43.118359371-06:00","closed_at":"2025-12-23T07:25:43.118359371-06:00"}
 {"id":"abide-kw5","title":"Add hard counting constraint primitives","description":"Add primitives that LLMs genuinely struggle with:\n- ExactCharacterCount(n): line must have exactly n characters\n- TotalVowelCount(n): entire poem must have exactly n vowels\n- WordLengthSequence: word N must have N letters\n- CrossLineArithmetic: line N word count = line N-1 vowel count\n- NoSharedLetters(line_a, line_b): two lines share no letters","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-09T07:28:54.728858497-06:00","updated_at":"2025-12-09T07:31:47.585951402-06:00","closed_at":"2025-12-09T07:31:47.585951402-06:00"}
+{"id":"abide-l12q","title":"GRPO parameter search experiment","description":"Systematic search for optimal max_tokens and repetition_penalty for Baguettotron","status":"open","priority":2,"issue_type":"epic","created_at":"2025-12-23T20:52:16.579768188-06:00","updated_at":"2025-12-23T20:52:16.579768188-06:00"}
 {"id":"abide-l5q","title":"Add Bop form (3 stanzas with repeated refrain)","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-07T19:04:26.733228377-06:00","updated_at":"2025-12-08T11:38:10.102057564-06:00","closed_at":"2025-12-08T11:38:10.102057564-06:00"}
 {"id":"abide-l9u","title":"Add variable syllable pattern to Limerick","description":"\n## Current State\nLimerick verifies: LineCount(5), RhymeScheme('AABBA')\nDescription says lines 1,2,5 should have 7-10 syllables; lines 3,4 should have 5-7\n\n## Missing Verification\nVariable syllable patterns NOT verified\nCurrently only checks line count and rhyme\n\n## Implementation\nAdd SyllablesPerLine with ranges:\n- Lines 1,2,5: NumericBound.between(7, 10)\n- Lines 3,4: NumericBound.between(5, 7)\n\n## Alternative\nCould use LineLengthRange constraint\n\n## Files\n- src/abide/forms/limerick.py\n","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T00:48:03.845234167-06:00","updated_at":"2025-12-11T00:56:19.815674787-06:00","closed_at":"2025-12-11T00:56:19.815674787-06:00"}
 {"id":"abide-l9um","title":"Audit scoring: FibonacciVerse","description":"Check verify() scoring - fix if too lenient","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T07:09:01.632114755-06:00","updated_at":"2025-12-23T07:25:30.563488149-06:00","closed_at":"2025-12-23T07:25:30.563488149-06:00"}
@@ -442,4 +446,5 @@
 {"id":"abide-zi42","title":"Audit scoring: Etheree","description":"Check verify() scoring - fix if too lenient","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T07:08:41.301551323-06:00","updated_at":"2025-12-23T07:22:05.011260147-06:00","closed_at":"2025-12-23T07:22:05.011260147-06:00"}
 {"id":"abide-zj2c","title":"Audit scoring: ShakespeareanSonnet","description":"Check verify() scoring - fix if too lenient","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-23T07:14:37.562263383-06:00","updated_at":"2025-12-23T07:20:37.303207429-06:00","closed_at":"2025-12-23T07:20:37.303207429-06:00"}
 {"id":"abide-zjah","title":"Audit SquareStanzas verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:55:17.154210412-06:00","updated_at":"2025-12-11T11:06:27.96879263-06:00","closed_at":"2025-12-11T11:06:27.96879263-06:00"}
+{"id":"abide-zuf1","title":"Analyze results and confirm best params","description":"Identify best params, repeat to confirm not stochastic","status":"open","priority":2,"issue_type":"task","created_at":"2025-12-23T20:52:34.70761466-06:00","updated_at":"2025-12-23T20:52:34.70761466-06:00","dependencies":[{"issue_id":"abide-zuf1","depends_on_id":"abide-8teh","type":"blocks","created_at":"2025-12-23T20:53:08.338463568-06:00","created_by":"darren"}]}
 {"id":"abide-zz33","title":"Audit ProsePoem verification","description":"","status":"closed","priority":2,"issue_type":"task","created_at":"2025-12-11T10:52:30.406558742-06:00","updated_at":"2025-12-11T11:05:19.340537215-06:00","closed_at":"2025-12-11T11:05:19.340537215-06:00"}

scripts/run_grpo.sh

@@ -6,7 +6,7 @@ set -e
 # (GPU order matters for NCCL - training must be on GPU 0)
 
 # Configuration
-MODEL="${ABIDE_MODEL:-google/gemma-3-270m-it}"
+MODEL="${ABIDE_MODEL:-allenai/OLMo-3-7B-Think-DPO}"
 PORT=8000
 VLLM_PID=""
 
@@ -38,16 +38,18 @@ sleep 2
 mkdir -p logs
 
 # Start vf-vllm on GPU 1
+# Using V1 engine (patched scheduler assertion)
 echo "Starting vf-vllm on GPU 1..."
 CUDA_VISIBLE_DEVICES=1 nohup /home/darren/miniconda3/bin/vf-vllm \
     --model "$MODEL" \
     --port $PORT \
-    --gpu-memory-utilization 0.7 \
+    --gpu-memory-utilization 0.92 \
     --tensor-parallel-size 1 \
-    --max-model-len 1024 \
+    --max-model-len 4096 \
     --trust-remote-code \
     --disable-log-stats \
     --enforce-eager \
+    --dtype bfloat16 \
     > logs/vllm.log 2>&1 &
 
 VLLM_PID=$!