Instructions to use ricdomolm/talkie-web-coder with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ricdomolm/talkie-web-coder with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ricdomolm/talkie-web-coder", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("ricdomolm/talkie-web-coder", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use ricdomolm/talkie-web-coder with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ricdomolm/talkie-web-coder"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ricdomolm/talkie-web-coder",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ricdomolm/talkie-web-coder

SGLang

How to use ricdomolm/talkie-web-coder with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "ricdomolm/talkie-web-coder" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ricdomolm/talkie-web-coder",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "ricdomolm/talkie-web-coder" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ricdomolm/talkie-web-coder",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ricdomolm/talkie-web-coder with Docker Model Runner:
```
docker model run hf.co/ricdomolm/talkie-web-coder
```

talkie-web-coder / README.md

ricdomolm

Update README.md

3834424 verified 21 days ago

preview code

raw

history blame contribute delete

2.45 kB

	---
	license: apache-2.0
	language:
	- en
	tags:
	- code
	- swe-bench
	- agentic
	- sft
	library_name: transformers
	pipeline_tag: text-generation
	datasets:
	- ricdomolm/mini-coder-trajs-400k
	base_model:
	- talkie-lm/talkie-web-13b-base
	---

	# talkie-web-coder

	13B model fine-tuned on agentic software-engineering trajectories from
	[SWE-smith](https://github.com/SWE-bench/SWE-smith), starting from the
	`talkie-web` base (same architecture as `talkie-1930` but pre-trained on
	web-style data). Tuned for the
	[mini-swe-agent](https://github.com/SWE-bench/mini-swe-agent) interaction
	format.

	## SWE-bench-Verified-Working-Harbor pass@1

	\| metric \| value \|
	\|---\|---\|
	\| pass@1 (n=3 independent eval runs) \| 5.75% ± 1.04 pp \|
	\| per-run resolved (out of 446) \| 31, 23, 23 \|

	Eval pipeline: vLLM (`--model-impl transformers --max-model-len 32768
	--dtype bfloat16`) → mini-swe-agent (`mini-extra swebench`, temperature 0.7,
	`max_tokens=4096`), graded with the swebench harness against
	`ricdomolm/SWE-bench_Verified-Working-Harbor`.

	## Training recipe

	\| \| \|
	\|---\|---\|
	\| Base model \| `talkie-web-13b-base` (chat-token reinitialised) \|
	\| Dataset \| `talkie-web-swe-100k-64k` (100k SWE-smith trajectories, packed at 64k) \|
	\| Trainer \| TRL `SFTTrainer` via `accelerate` (8× A100) \|
	\| Optimizer \| `adamw_torch_fused`, β=(0.9, 0.95), ε=1e-8 \|
	\| LR \| 2e-5, `cosine_with_min_lr`, warmup 3% \|
	\| Precision \| bf16 \|
	\| Weight decay \| 0.1 \|
	\| Max grad norm \| 30 \|
	\| Max length \| 65,536 \|
	\| Packing \| `bfd` + padding-free \|
	\| Loss \| `completion_only_loss=1` (loss only on assistant tokens) \|
	\| Steps \| 2,016 (this is ckpt-2000) \|

	## Usage

	This model uses custom modeling code (`modeling_talkie.py`,
	`configuration_talkie.py`). Load with `trust_remote_code=True`:

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer

	model = AutoModelForCausalLM.from_pretrained(
	"ricdomolm/talkie-web-coder",
	trust_remote_code=True,
	torch_dtype="bfloat16",
	)
	tokenizer = AutoTokenizer.from_pretrained("ricdomolm/talkie-web-coder")
	```

	For agentic eval, serve with vLLM and drive with mini-swe-agent:

	```bash
	vllm serve ricdomolm/talkie-web-coder \
	--model-impl transformers --max-model-len 32768 --dtype bfloat16
	```

	## Companion model

	[`ricdomolm/talkie-1930-coder`](https://huggingface.co/ricdomolm/talkie-1930-coder)
	— same recipe, same SFT data, but starting from a different base model.
	Reaches 4.48% ± 0.69 pp on the same eval (n=5).