OELLM 9B Anneal-300B Tool-Use SFT

An experimental full-parameter supervised fine-tune of Neonkraft/oellm-9b-256k-theta64m-prelude-anneal300b-instruct-sft for function calling. The training mix combines English tool-use conversations with general instruction replay. This is an SFT checkpoint; it has not received preference optimization or reinforcement learning. It is separate from the Qwen3.5-4B-EU-Tool checkpoint, which was a format reference and was not a training parent.

Status: Training completed on LUMI on 7 October 2026. Paired held-out tool-call evaluation completed on 9 October 2026. The model calls available tools accurately on this test but overcalls on unrelated requests; see the evaluation table below.

Model and format

The architecture is Qwen3ForCausalLM with 9,101,947,904 parameters and the OpenEuroLLM/Qwen3-style ChatML token mapping inherited from the parent. Available tools are described in <tools>...</tools> in the system turn; the assistant emits <tool_call>{"name": ..., "arguments": {...}}</tool_call>. Tool replies are represented as <tool_response>...</tool_response> in a user turn. An empty <think>\n\n</think>\n\n block precedes assistant responses after the latest user query. <|im_end|> (token ID 4) terminates assistant turns. The packaged chat_template.jinja implements this format.

Context and RoPE caveat: The parent advertises a 256K context and RoPE theta 64,000,000. This continuation was trained with Transformers 4.57.3, which resolved the parent configuration to an effective RoPE theta of 10,000 because the 64M value was stored in rope_parameters while that version's Qwen3 rotary implementation read rope_theta. The exported configuration explicitly sets both fields to 10,000 to preserve the behavior used during training. Training sequences were at most 4,096 tokens. Long-context performance, including 256K tool use, has not been validated. The max_position_embeddings architecture field remains 262,144; it is not a validated context-length claim for this checkpoint.

Training used BF16 computation. The saved full-parameter weights are FP32 (about 36 GB), so plan memory accordingly or explicitly load in BF16 for inference.

Training data

The tool-use dataset has 55,077 source rows and is derived from glaiveai/glaive-function-calling-v2 (50,000 rows), Team-ACE/ToolACE (3,184), and NousResearch/hermes-function-calling-v1 (1,893). Its dataset card describes all three as Apache-2.0 and documents their normalization into <tool_call> JSON. No examples from its grpo configuration were used.

Preparation trimmed trailing non-assistant turns, filtered 14 overlength tool rows, and removed 16,473 duplicate tool prompts keyed on both the system tool schema and the first user request. A deterministic 5% prompt split held out 1,948 tool conversations (about 1,048,852 rendered tokens). Dolci replay was selected from the parent training artifact until it matched the retained tool training tokens. There were only 10 no-call conversations in the entire tool-use source after normalization. This is a major limitation for learning when to abstain from calling a tool.

The counts and checksums above come from the saved training_data_manifest.json. They were computed during data construction with the earlier chat template; the final tool-use template adds an empty think block to intermediate tool calls, so these token counts are approximate for the actual training serialization.

Training procedure

The parent revision was 85bf18fb4f0bee6ac6270f06b1d1c6b3be200f31. The training data manifest SHA-256 was caf837e99a93e2bcc9b71509980e60eba8895184216c541bc52548f5a92a3ac0. The train Parquet SHA-256 was 01cbe9ecd71b7ef6f3b6c177ad06cc9692ff8489811eb13dc83b07a673d1f87f. The repository state was based on BirgerMoell/oellm-reasoning-training commit 7b9f05a66bdfd6ca5b0d5e9c88cdabcc69098ca2 plus local uncommitted tool-use changes. LUMI Slurm job: 22546173.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "birgermoell/oellm-9b-anneal300b-tooluse-sft"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id, torch_dtype=torch.bfloat16, device_map="auto"
).eval()

tools = [{"type": "function", "function": {
    "name": "get_weather",
    "description": "Get weather for a city",
    "parameters": {"type": "object", "properties": {
        "city": {"type": "string"}}, "required": ["city"]}
}}]
messages = [{"role": "user", "content": "What is the weather in Helsinki?"}]
inputs = tokenizer.apply_chat_template(
    messages, tools=tools, add_generation_prompt=True,
    enable_thinking=False, return_dict=True, return_tensors="pt",
).to(model.device)
inputs.pop("token_type_ids", None)  # needed with some Transformers versions
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=192, do_sample=False)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[-1]:],
                       skip_special_tokens=False))

The dataset embedded tool signatures directly in system messages. Supplying tools= as above renders the same tool wrapper through the packaged template. Execute any returned tool call in your own application, then append the result as a tool message and generate again. The model weights do not execute tools.

With Transformers 4.57.3, the tokenizer's saved metadata may trigger a TypeError in AutoTokenizer. A compatible fallback is to load tokenizer.json with PreTrainedTokenizerFast and set its chat_template from chat_template.jinja (both files are in this repository). The training and evaluation code use that fallback.

Evaluation and limitations

The paired held-out evaluation compares this checkpoint with its direct parent using greedy decoding and at most 192 new tokens. It uses 100 callable prompts from the held-out tool split and 100 untouched Dolci instruction prompts paired with an irrelevant calculate_tip tool. The latter is a synthetic relevance probe; the tool dataset itself contained too few no-call conversations for a balanced held-out sample.

The continuation learned call syntax and argument selection on this test, but it called the irrelevant available tool on 73 of 100 unrelated prompts. The parent produced no well-formed calls on the callable prompts. The result supports improving no-call/relevance training before relying on this model in an agent that exposes optional tools. Training loss alone did not reveal this tradeoff.

The test is narrower than full BFCL: it has only 200 examples, the no-call class uses an artificial unrelated tool, and it does not measure multi-call planning, tool execution, long-context behavior, or real-world tool choice. The tokenizer loaders used for the parent and candidate produced identical input IDs on all 200 evaluation prompts despite a Transformers warning about the candidate tokenizer regex. LUMI evaluation job 22605103 completed in 44 minutes 23 seconds; the machine-readable results are in the training workspace at artifacts/eval/tooluse-anneal300b-v2/paired-heldout.json.

The tool-use training data is English-only. The model can produce malformed arguments, select the wrong function, or call a tool when none is needed. No safety or multilingual tool-use benchmark has been run for this checkpoint.

License and attribution

The parent model and tool-use dataset are Apache-2.0. The Dolci instruction replay is ODC-By-1.0; retain its attribution and review the upstream dataset card and responsible-use terms. Training was performed on LUMI, hosted by CSC in Finland, for OpenEuroLLM research.