Fix bounded DeepSeek reasoning and answer generation

README.md

@@ -43,17 +43,19 @@ python script.py --input path/to/test.csv --output submission.csv
 python script.py --self-test
 ```
 
-The output file is initialized before model loading and atomically rewritten
-after every completed row. If evaluation reaches its time limit, predictions
-already completed remain in a valid submission file.
+After the model loads successfully, the output file is initialized and then
+atomically rewritten after every completed row. If evaluation reaches its time
+limit, predictions already completed remain in a valid submission file.
 
 ## DeepSeek reasoning and generation
 
 DeepSeek's solver instructions, retrieved context, and current problem are
 placed in one user message rather than a separate system message. The assistant
 prompt is prefixed with `<think>` to engage the distilled reasoning behavior.
-Completed reasoning content is kept separately from the visible response; only
-the strict `FINAL ANSWERS:` block is serialized into `pred`.
+Reasoning stops when `</think>` appears or at its configured cap. The runtime
+then starts a separate `FINAL ANSWERS:` stage, guaranteeing that reasoning
+cannot consume the answer budget. Only that answer block is serialized into
+`pred`.
 
 Defaults follow DeepSeek's recommended sampling values while retaining a bounded
 challenge-time output:
@@ -61,8 +63,9 @@ challenge-time output:
 - `temperature=0.6`
 - `top_p=0.95`
 - sampling enabled
-- `max_new_tokens=2048`
-- at most two attempts, with the second used only after a parser failure
+- `max_reasoning_tokens=4096`
+- `max_answer_tokens=1024`
+- one answer-only retry with `answer_retry_tokens=512`
 
 The model context is capped at 32,768 tokens. The script loads AWQ directly in
 FP16, enables the KV cache, and applies an inference-only last-token `lm_head`
@@ -78,8 +81,9 @@ indexed.
 
 Settings are in `rag_resources/config.json`. Environment overrides are available
 for `IOL_TOP_METHODS`, `IOL_TOP_EXAMPLES`, `IOL_CHAR_TFIDF_WEIGHT`,
-`IOL_RAG_MAX_CHARS`, `IOL_MAX_NEW_TOKENS`,
-`IOL_EXPLANATION_MAX_NEW_TOKENS`, `IOL_MAX_ATTEMPTS`,
+`IOL_RAG_MAX_CHARS`, `IOL_MAX_REASONING_TOKENS`,
+`IOL_MAX_ANSWER_TOKENS`, `IOL_ANSWER_RETRY_TOKENS`,
+`IOL_EXPLANATION_MAX_NEW_TOKENS`,
 `IOL_ENABLE_EXPLANATIONS`, and `IOL_SEED`.
 
 ## Sources and licenses

rag_resources/config.json

@@ -4,9 +4,10 @@
   "char_tfidf_weight": 3.0,
   "rag_max_chars": 12000,
   "model_context_tokens": 32768,
-  "max_new_tokens": 2048,
+  "max_reasoning_tokens": 4096,
+  "max_answer_tokens": 1024,
+  "answer_retry_tokens": 512,
   "explanation_max_new_tokens": 400,
-  "max_attempts": 2,
   "enable_explanations": false,
   "temperature": 0.6,
   "top_p": 0.95,

rag_resources/system_prompt.txt

@@ -17,6 +17,6 @@ Common task types and what to return:
 - number to text: the number written in words in the requested language;
 - any other type: exactly what the instruction requests, with no extra material in the answer.
 
-Reason step by step, but keep the analysis focused and always reserve enough of the 2048-token output budget for the final answer. Finish with a line that says exactly:
+Reason step by step, but keep the analysis focused. The runtime reserves a separate answer stage, so complete the analysis as soon as you have a consistent solution. Finish with a line that says exactly:
 FINAL ANSWERS:
 Below it, put one bare answer per requested item or blank, in order. Do not number the lines. Do not add quotes, labels, explanations, or commentary inside that final block.