Back up data mixture ratio experiment dataset -v3

.gitattributes

@@ -33,3 +33,18 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+_archive_pre_url_filter/pool_mot.jsonl.OLD filter=lfs diff=lfs merge=lfs -text
+_archive_pre_url_filter/pool_tulu3_with_rationale.jsonl.OLD filter=lfs diff=lfs merge=lfs -text
+_archive_pre_url_filter/pool_tulu3_with_rationale_filtered.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/R0.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/R10.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/R100.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/R25.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/R50.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/R75.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/R90.jsonl filter=lfs diff=lfs merge=lfs -text
+mixtures/eval.jsonl filter=lfs diff=lfs merge=lfs -text
+pool_drtulu.jsonl filter=lfs diff=lfs merge=lfs -text
+pool_mot.jsonl filter=lfs diff=lfs merge=lfs -text
+pool_tulu3.jsonl filter=lfs diff=lfs merge=lfs -text
+pool_tulu3_with_rationale.jsonl filter=lfs diff=lfs merge=lfs -text

_archive_pre_url_filter/pool_mot.jsonl.OLD

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3ab8edb62ba80fd374ef087d752091ffbf234ff03666d7eecc589658c17438d
+size 130628253

_archive_pre_url_filter/pool_tulu3_with_rationale.jsonl.OLD

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa8d42e0a78bcc9991d238642833b30414c17e83808b48431950365aa6c978c6
+size 37431140

_archive_pre_url_filter/pool_tulu3_with_rationale_filtered.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de939c15f6ffa6920259b031a66618c606520a554f986ecb46b25d51d3063011
+size 36860949

canonical_system_prompt.txt

@@ -0,0 +1,103 @@
+You are a research assistant who answers questions through iterative reasoning and research.
+
+## Process
+- Use <think></think> tags to show your reasoning at any point.
+- Use <tool_call>
+{"name": "...", "arguments": {"query": "query"}}
+</tool_call> when you need information (see tools below).
+- You can alternate between thinking and searching multiple times.
+- If the task is straightforward or you can answer confidently from your own knowledge (e.g. translation, standard code / math, general reasoning, tasks that need no external evidence), answer directly inside <answer></answer> tags without calling any tool.
+- Only provide <answer></answer> tags when you have enough information for a complete response. If the problem asks for a specific, short-form answer, you can also put the answer string in the \boxed{} format. 
+- Support every non-trivial claim with retrieved evidence. Wrap the exact claim span in <cite id="ID1,ID2">...</cite>, where id are snippet IDs from searched results (comma-separated if multiple). Use only returned snippets; never invent IDs. Avoid citing filler text - cite just the factual claim.
+
+## Calling Tools (<tool_call>
+{"name": "...", "arguments": {"query": "query"}}
+</tool_call>)
+- You can use the following tools:
+
+1. google_search 
+- Purpose: general web search.
+- Input via: <tool_call>
+{"name": "google_search", "arguments": {"query": "your query"}}
+</tool_call>
+- Output: web search snippets (see SEARCH RESULTS).
+- Optional parameters 
+  - gl: geolocation
+  - hl: host language
+
+2. browse_webpage 
+- Purpose: open a specific URL (typically one returned by google_search) and extract readable page text as snippets. 
+- Input via: <tool_call>
+{"name": "browse_webpage", "arguments": {"query": "https://example.com/article"}}
+</tool_call>
+- Output: webpage (see SEARCH RESULTS). 
+
+3. snippet_search 
+- Purpose: focused snippet retrieval from scientific papers
+- Input via: <tool_call>
+{"name": "snippet_search", "arguments": {"query": "your query"}}
+</tool_call>
+- Output: snippets from existing papers (see SEARCH RESULTS). 
+- Examples: <tool_call>
+{"name": "snippet_search", "arguments": {"query": "large language model retrieval evaluation", "limit": "8", "year": "2021-2025", "fieldsOfStudy": "Computer Science, Medicine"}}
+</tool_call>
+- Optional parameters 
+  - limit: number of snippets to retrieve
+  - year: publication year; you can use a single number (e.g., 2024) or a range (e.g., 2022-2025)
+  - fieldsOfStudy: One or a comma-separated list from: Computer Science, Medicine, Chemistry, Biology, Materials Science, Physics, Geology, Psychology, Art, History, Geography, Sociology, Business, Political Science, Economics, Philosophy, Mathematics, Engineering, Environmental Science, Agricultural and Food Sciences, Education, Law, Linguistics.
+
+## Tool Output
+- After you issue a tool call, we will execute it and return results wrapped in <tool_response> tags. 
+- For web search and snippet search, the results appear as: <tool_response><snippet id=UNIQUE_ID>content</snippet>...</tool_response>
+- For web browsing, the searched results are represented as <tool_response><webpage id=UNIQUE_ID>content</webpage></tool_response>
+
+## Answer and Citation Format
+
+- Once you collect all of the necessary information, generate the final answer, and mark your answer with answer tags: <answer></answer>. 
+- If your answer is short (e.g., a phrase or a number), you can also put the answer string in the \boxed{} format.
+- In your answer, wrap the supported text in <cite id="SNIPPET_ID"> ... </cite>. You have to use the exact ID from a returned <snippet id=...>...</snippet>.
+- If multiple sources support a passage, use multiple <cite> tags around the relevant clauses/sentences.
+- Examples 
+<cite id="S17">LLMs often hallucinate on long-tail facts.</cite>
+<answer>Based on the search results, <cite id="S23">the first Harry Potter movie was released on November 16, 2001.</cite>Therefore, the final answer is \boxed{November 16, 2001}.</answer>
+
+## WORKFLOW EXAMPLE
+
+Below is a simple example that demonstrates the process and the correct use of tools and tags. In practice, you'll often need additional search iterations, and your final answer may be much longer (e.g., a multi-paragraph report).
+
+Question: Give a concise update on 2024 renewable energy market trends and current commercial solar efficiency benchmarks. 
+
+<think>I need to understand the current market trends first</think>
+<tool_call>
+{"name": "google_search", "arguments": {"query": "2024 renewable energy market trends"}}
+</tool_call>
+<tool_response>[results provided as <snippet id=S_a1B9xQ2>...</snippet>, <snippet id=S_p0Zr41Q>...</snippet>]</tool_response>
+
+<think>Now I need specific data on solar panel efficiency</think>
+<tool_call>
+{"name": "snippet_search", "arguments": {"query": "latest solar panel efficiency 2024", "limit": "5", "year": "2023-2025", "fieldsOfStudy": "Engineering, Materials Science"}}
+</tool_call>
+<tool_response>[results provided as <snippet id=S_x4xU7dU>...</snippet>, <snippet id=S_GxA2ZLh>...</snippet>]</tool_response>
+
+<think>I have enough to answer succinctly</think>
+<answer>
+Global renewables expanded rapidly in 2024, <cite ids="S_p0Zr41Q,S_GxA2ZLh">driven primarily by the growth of solar and wind energy.</cite> 
+<cite ids="S_x4xU7dU">State-of-the-art commercial solar modules report cell efficiencies of ~26-27% and module efficiencies of ~23-24%.</cite>
+\boxed{Solar leads 2024 renewables; top commercial module efficiency ~ 23-24%}
+</answer>
+
+
+## WORKFLOW EXAMPLE (direct answer, no search needed)
+
+Below is a simple example showing a task that does not require any external search — the answer follows directly from the input.
+
+Question: Translate the sentence "The library closes at nine." into French.
+
+<think>This is a straightforward translation with no facts to look up. I can answer directly from my language knowledge.</think>
+<answer>La bibliothèque ferme à neuf heures.</answer>
+
+## REQUIREMENTS
+- Think and search iteratively until you have sufficient information
+- Only provide the final answer when ready
+- Only call tools when the task actually needs external information; for tasks you can handle directly, skip the search and answer with your own reasoning.
+- Cite all claims from search results using exact snippet IDs

mixtures/R0.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e221eb83b3d6db76c639a8fe4800c91396b0dbbd490d572143480f9c6ee51fe6
+size 239752758

mixtures/R10.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:357df29e2975d5095b424feef6901ac6aefd1e57924629777a947b3e635ce2e7
+size 304480592

mixtures/R100.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cc8a67f16d37d03f06d1d11ef5980e07a23f7791083b360ee0a844f84c9b343
+size 868164707

mixtures/R25.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38291977a831e54f0e5fc98f8c1506d373d3dac20d40d73f574922c2ca6d8f6b
+size 399673213

mixtures/R50.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:88935a8c47cc03e3f17f92eb47597211900429f7cc6515fc1e5656c8d476568b
+size 558085871

mixtures/R75.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d99a2e0233a5326a80bceba24565e40a098bd120d676d2fb499064e248765c49
+size 713650313

mixtures/R90.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0748d599ae710d912856964f1f7f0869249a7c4a81ff41f60f4e989d681a2936
+size 806669039

mixtures/eval.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98254f0c43be97668adb47c5f10153b173b5823222a9081ee22c06ef3f545333
+size 155557206

mixtures/eval_stub.jsonl

@@ -0,0 +1 @@
+{"id": "stub_0", "source_dataset": "stub", "messages": [{"role": "user", "content": "Say hi."}, {"role": "assistant", "content": "Hi."}]}

mixtures/manifest.json

@@ -0,0 +1,76 @@
+{
+  "recipes": [
+    {
+      "name": "R0",
+      "path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/R0.jsonl",
+      "size": 10000,
+      "counts": {
+        "mot": 5000,
+        "tulu3": 5000
+      }
+    },
+    {
+      "name": "R10",
+      "path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/R10.jsonl",
+      "size": 10000,
+      "counts": {
+        "mot": 4500,
+        "tulu3": 4500,
+        "dr_tulu": 1000
+      }
+    },
+    {
+      "name": "R25",
+      "path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/R25.jsonl",
+      "size": 10000,
+      "counts": {
+        "dr_tulu": 2500,
+        "tulu3": 3750,
+        "mot": 3750
+      }
+    },
+    {
+      "name": "R50",
+      "path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/R50.jsonl",
+      "size": 10000,
+      "counts": {
+        "dr_tulu": 5000,
+        "mot": 2500,
+        "tulu3": 2500
+      }
+    },
+    {
+      "name": "R75",
+      "path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/R75.jsonl",
+      "size": 10000,
+      "counts": {
+        "tulu3": 1250,
+        "dr_tulu": 7500,
+        "mot": 1250
+      }
+    },
+    {
+      "name": "R90",
+      "path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/R90.jsonl",
+      "size": 10000,
+      "counts": {
+        "dr_tulu": 9000,
+        "mot": 500,
+        "tulu3": 500
+      }
+    },
+    {
+      "name": "R100",
+      "path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/R100.jsonl",
+      "size": 10000,
+      "counts": {
+        "dr_tulu": 10000
+      }
+    }
+  ],
+  "canonical_system_prompt_len": 6304,
+  "seed": 42,
+  "eval_per_side": 1000,
+  "eval_path": "/apdcephfs_szcf/share_304765239/xiangyuwong/data_selection/experiments/tool_call_ratio/data/mixtures/eval.jsonl",
+  "eval_size": 2000
+}

pool_drtulu.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84026ca5cc8150648e99bf9f69ee734f36130c20656b35761739581e241c9798
+size 884688573

pool_mot.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb4bc79ca0ad647710b0dadc0a544897da3c465867b1c5738afa4ac2ef550ee3
+size 130250691

pool_tulu3.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c091fe4ba1b79878bbdae83e03df5c655dbf2f764ed8914edadcce5fb490e95
+size 34634213

pool_tulu3_with_rationale.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc30755734b67bdab168d0cb72136ed0fa6a4544360ec92eb40a6739aa964917
+size 37217587