Instructions to use hunterbown/trainless-toolcall-2b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use hunterbown/trainless-toolcall-2b with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B") model = PeftModel.from_pretrained(base_model, "hunterbown/trainless-toolcall-2b") - Notebooks
- Google Colab
- Kaggle
Add the result on untouched categories: 1.6 points below the base where most requests fit no tool
eb2e18d verified Download trainless_model.json from hunterbown/trainless-toolcall-2b: direct link, hf CLI and curl.
- Browser
- Download file 8.35 kB
-
https://huggingface.co/hunterbown/trainless-toolcall-2b/resolve/main/trainless_model.json
- Command line
-
hf download hf://hunterbown/trainless-toolcall-2b/trainless_model.json
-
curl -L -o trainless_model.json https://huggingface.co/hunterbown/trainless-toolcall-2b/resolve/main/trainless_model.json
8.35 kB
| { | |
| "format": "trainless-peft/1", | |
| "base_model": "Qwen/Qwen3.5-2B", | |
| "base_weights_sha256": "aa33250c4fc64891ddfaba3a314fd9542ea371843c387178b425fbcc5ed680b1", | |
| "auto_class": "AutoModelForImageTextToText", | |
| "runtime": "transformers-peft", | |
| "method": "forward-only analytic weight editing (project_out on one MLP writer, strength 4)", | |
| "quantized_runtime_tested": false, | |
| "edit": { | |
| "tensor": "model.language_model.layers.10.mlp.down_proj.weight", | |
| "op": "project_out", | |
| "space": "rows", | |
| "strength": 4.0, | |
| "direction_sha256": "817ceab1340a687015a83c95eeeec49017079e99113a4166958724f3f7d48952", | |
| "trainless_session": "ss_daf42ea7528e", | |
| "state_id": "182dc3991ca48142", | |
| "edited_tensor_sha256": "d5e4e6cdbc7c555e023198ab79a85c614b0de064f9416fab2514c3274dd28d26", | |
| "fraction_of_editable_weights_changed": 0.006237529234172905 | |
| }, | |
| "refusal_check": { | |
| "base_rate": 0.984375, | |
| "benchmark": "AdvBench harmful_behaviors (first 64), greedy, thinking off, 256 tokens, refusal-phrase grader", | |
| "drop": 0.0, | |
| "edited_rate": 0.984375, | |
| "limit": 0.2, | |
| "n": 64, | |
| "passed": true, | |
| "state_id": "182dc3991ca48142", | |
| "ts": "2026-11-11T00:09:45.234740882+00:00" | |
| }, | |
| "evaluation": { | |
| "benchmark": "BFCL v4 (ShishirPatil/gorilla @ 6ea57973c7a6097fd7c5915698c54c17c5b1b6c8), categories simple_python,multiple,parallel,parallel_multiple,irrelevance,live_simple", | |
| "split": "test: sha256(category/id) mod 4 in {2,3}; construction split 0 built the direction, dev split 1 chose layer and strength", | |
| "n": 703, | |
| "engine": "vllm 0.31.0", | |
| "decoding": "greedy", | |
| "thinking": false, | |
| "max_new_tokens": 1024, | |
| "stop_token_ids": [ | |
| 248044, | |
| 248046 | |
| ], | |
| "grading": "BFCL AST checker, unchanged; irrelevance is correct when no call is made", | |
| "strong_prompt": "You are a precise function-calling assistant. Rules:\n1. If a provided function can answer the request, call it. Use the exact function name.\n2. Copy argument values exactly from the user's request; do not invent, round or rename them.\n3. Include every required parameter. Include optional parameters only when the user gives a value.\n4. Use the parameter's declared type (integers as integers, lists as JSON lists, booleans as true/false).\n5. If the request needs several calls, emit one <tool_call> block per call.\n6. If no provided function fits the request, do not call any function; answer in plain text.", | |
| "adapter_sha256": { | |
| "adapter_config.json": "70c7c9a027f88e37d92d66c7058a43708a481ef7486791b19f5ee8e402712346", | |
| "adapter_model.safetensors": "32a8d4336a298a3329c033be4c4c38c9d15afabd4a3eee4bbafd76b9db9ec808", | |
| "export_report.json": "32ce0bdb34d75aa5ba64ff97890fe1df382106b331e34e27bd0f6e70ad4831ed" | |
| }, | |
| "arms": { | |
| "normal": { | |
| "correct": 526, | |
| "n": 703, | |
| "relevant_cases_with_no_call": 84, | |
| "irrelevant_cases_with_a_call": 18, | |
| "capped": 7 | |
| }, | |
| "strong": { | |
| "correct": 544, | |
| "n": 703, | |
| "relevant_cases_with_no_call": 10, | |
| "irrelevant_cases_with_a_call": 58, | |
| "capped": 3 | |
| }, | |
| "adapter": { | |
| "correct": 552, | |
| "n": 703, | |
| "relevant_cases_with_no_call": 46, | |
| "irrelevant_cases_with_a_call": 23, | |
| "capped": 7 | |
| } | |
| }, | |
| "by_category": { | |
| "irrelevance": { | |
| "normal": 89, | |
| "strong": 49, | |
| "adapter": 84, | |
| "n": 107 | |
| }, | |
| "live_simple": { | |
| "normal": 90, | |
| "strong": 92, | |
| "adapter": 85, | |
| "n": 125 | |
| }, | |
| "multiple": { | |
| "normal": 88, | |
| "strong": 86, | |
| "adapter": 88, | |
| "n": 102 | |
| }, | |
| "parallel": { | |
| "normal": 40, | |
| "strong": 72, | |
| "adapter": 53, | |
| "n": 86 | |
| }, | |
| "parallel_multiple": { | |
| "normal": 70, | |
| "strong": 74, | |
| "adapter": 75, | |
| "n": 89 | |
| }, | |
| "simple_python": { | |
| "normal": 149, | |
| "strong": 171, | |
| "adapter": 167, | |
| "n": 194 | |
| } | |
| }, | |
| "paired": { | |
| "adapter_minus_normal": { | |
| "accuracy_delta": 0.037, | |
| "bootstrap95": [ | |
| 0.0171, | |
| 0.0569 | |
| ], | |
| "first_wrong_second_right": 38, | |
| "first_right_second_wrong": 12 | |
| }, | |
| "adapter_minus_strong_prompt": { | |
| "accuracy_delta": 0.0114, | |
| "bootstrap95": [ | |
| -0.0171, | |
| 0.0413 | |
| ], | |
| "first_wrong_second_right": 58, | |
| "first_right_second_wrong": 50 | |
| }, | |
| "note": "paired case bootstrap, 10,000 resamples, seed 0; greedy decoding, so no sampling-seed variance, but other runtimes can differ" | |
| }, | |
| "controls": { | |
| "note": "scripted HF Transformers run of the bf16 edit on the same 703 cases, not this adapter in vLLM", | |
| "edit": 554, | |
| "normal": 524, | |
| "random_direction_same_layer_and_strength": [ | |
| 510, | |
| 522, | |
| 522 | |
| ] | |
| }, | |
| "what_changed": { | |
| "relevant_cases_base_already_called": { | |
| "n": 512, | |
| "base_correct": 437, | |
| "adapter_correct": 433 | |
| }, | |
| "relevant_cases_base_answered_in_text": { | |
| "n": 84, | |
| "base_correct": 0, | |
| "adapter_correct": 35 | |
| }, | |
| "fixes_where_base_made_no_call": "35 of 38" | |
| } | |
| }, | |
| "off_target": { | |
| "note": "vLLM, no tool definitions, thinking off, greedy, 4,096-token cap; base vs this adapter", | |
| "results": { | |
| "gsm8k_test": { | |
| "n": 500, | |
| "base_correct": 407, | |
| "adapter_correct": 399, | |
| "paired_diff_points": -1.6, | |
| "bootstrap95": [ | |
| -4.4, | |
| 1.2 | |
| ], | |
| "fixes": 20, | |
| "regressions": 28, | |
| "base_capped": 19, | |
| "adapter_capped": 19, | |
| "base_outputs_with_tool_call": 0, | |
| "adapter_outputs_with_tool_call": 0, | |
| "identical_outputs": 16 | |
| }, | |
| "math500": { | |
| "n": 200, | |
| "base_correct": 143, | |
| "adapter_correct": 144, | |
| "paired_diff_points": 0.5, | |
| "bootstrap95": [ | |
| -4.5, | |
| 5.5 | |
| ], | |
| "fixes": 13, | |
| "regressions": 12, | |
| "base_capped": 45, | |
| "adapter_capped": 44, | |
| "base_outputs_with_tool_call": 0, | |
| "adapter_outputs_with_tool_call": 0, | |
| "identical_outputs": 1 | |
| } | |
| } | |
| }, | |
| "made_by": { | |
| "agent": "Claude Fable 5.2 (claude-fable-5-2) in Claude Code", | |
| "human": "Hunter Bown set the goal and approved each stage and the release", | |
| "note": "the agent designed and ran the direction, screens, tests and export through scripts and the Trainless platform" | |
| }, | |
| "untouched_categories": { | |
| "note": "BFCL live_irrelevance, live_multiple, live_parallel, live_parallel_multiple; all cases; never used to build or test this adapter; vLLM, greedy, thinking off, 1,024-token cap", | |
| "n": 1977, | |
| "arms": { | |
| "normal": { | |
| "correct": 1397, | |
| "n": 1977, | |
| "relevant_no_call": 21, | |
| "irrelevant_calls": 273, | |
| "by_category": { | |
| "live_irrelevance": 611, | |
| "live_multiple": 766, | |
| "live_parallel": 10, | |
| "live_parallel_multiple": 10 | |
| } | |
| }, | |
| "strong": { | |
| "correct": 1292, | |
| "n": 1977, | |
| "relevant_no_call": 10, | |
| "irrelevant_calls": 371, | |
| "by_category": { | |
| "live_irrelevance": 513, | |
| "live_multiple": 755, | |
| "live_parallel": 12, | |
| "live_parallel_multiple": 12 | |
| } | |
| }, | |
| "a:published": { | |
| "correct": 1365, | |
| "n": 1977, | |
| "relevant_no_call": 16, | |
| "irrelevant_calls": 300, | |
| "by_category": { | |
| "live_irrelevance": 584, | |
| "live_multiple": 760, | |
| "live_parallel": 10, | |
| "live_parallel_multiple": 11 | |
| } | |
| } | |
| }, | |
| "paired": { | |
| "strong - normal": { | |
| "diff_points": -5.31, | |
| "ci95": [ | |
| -6.83, | |
| -3.79 | |
| ], | |
| "fixes": 69, | |
| "regressions": 174 | |
| }, | |
| "a:published - normal": { | |
| "diff_points": -1.62, | |
| "ci95": [ | |
| -2.43, | |
| -0.81 | |
| ], | |
| "fixes": 18, | |
| "regressions": 50 | |
| }, | |
| "a:published - strong": { | |
| "diff_points": 3.69, | |
| "ci95": [ | |
| 2.18, | |
| 5.26 | |
| ], | |
| "fixes": 155, | |
| "regressions": 82 | |
| } | |
| } | |
| } | |
| } |